SlideShare une entreprise Scribd logo
1  sur  22
Télécharger pour lire hors ligne
Analysis of Data-Based
Methods for Approximating
Fisher Information in the
Scalar Case
Shenghan Guo
Applied Math & Statistics, Johns Hopkins University
Bulletin:
Conclusion
Problem Setting:
Random variables with density function p(z|θ),
where the unknown parameter θ is scalar. Fisher information
is the measurement of information amount carried by z
regarding θ. 2
(θ) (θ| )nF E g =  z
[ ](θ) (θ| ) ,nF E H=− z
where
log ( |θ)
(θ| )
θ
p
g
∂
=
∂
z
z
and 2
2
log ( |θ)
(θ| )
θ
p
H
∂
=
∂
z
z
1 2[ , ,..., ]T
nZ Z Z=z
Applications:
Variance Approximation for Maximum
Likelihood Estimation (MLE)
Asymptotic normality: For some common form of
estimate , denote the true value of the parameter
by . Then under modest conditions,
ˆθn
A classical use of Fisher Information
θ*
1ˆ(θθ*) (0, (θ*))
dist
nn N F−
− →
A Fact :
It is difficult to get the true value of Fisher Information in
practice because direct calculation could be difficult, or
we don’t have the explicit forms of the gradient or the
Hessian. Thus approximation is needed.
Typical situations where approximation is
needed:
• State-space models with nuisance parameters (Spall,
J. C. and Garner, J. P. 1990)
• Non-i.i.d. projectile measurements and quantiles (Spall,
J. C. and Maryak, J. L. 1992)
• Inverse heat condition problems (Favennec, Y. 2007)
• The unavailability of the true Fisher information in
many practical cases makes a useful approximation
method for this quantity important.
• No prior studies have compared the approximation
methods for Fisher information to decide which one is
more accurate.
• The convention of utilizing the certain approximation
method (e.g. 2nd
-derivative approximation to estimate
Fisher information in the scalar case) needs
theoretical support.
Motivation:
Two Methods to Approximate Fisher
Information Number:
I. Sums of Products of Derivatives:
II. Negative Sum of 2nd
Derivative:
2
1
1
(θ) (θ| )
n
n i i
i
G g z
n =
= ∑
1
1
(θ) (θ| )
n
n i i
i
H H z
n =
= − ∑
The reality: Mostly people tend to use the 2nd
derivative method instead of the product-of-
derivatives method, which is more of a convention
without valid theoretical support.
My Way to Solve This Question:
By the Central Limit Theorem:
Let and , for independent and
identically distributed random variables:
( ) ( )dist 2
(θ) (θ) 0,var ( )n in G F n N g z − →  
( ) [ ]( )dist
(θ) (θ) 0,var ( )n in H F n N H z− →
2 2
( ) (θ| )g g• = • ( ) (θ| )H H• = •
nG
nH
We wonder which one of the estimates and
is more accurate !
For independent and non-identically distributed
random variables:
( ) ( )dist
(θ) (θ) 0,n gn G F n N V− →
( ) ( )dist
(θ) (θ) 0,n Hn H F n N V−  →
1 2
1
lim var ( )
n
g i iin
V n g z−
=→∞
 =  ∑where
and [ ]1
1
lim var ( )
n
H i iin
V n H z−
=→∞
= ∑
The Challenge:
Calculation of the asymptotic variances in
above CLT is infeasible because we cannot
compute the variances analytically !
How to solve this problem?
----Estimation by Taylor expansion
Assuming that our density function is at least twice-
differentiable with respect to θ, we use the second-
order Taylor expansion to approximate the asymptotic
variances.
2
var ( )ig z 
 
Take expectation of the Taylor expansion
for the product of g(zi)
[ ]var ( )iH z
Take expectation of the Taylor expansion
for H(zi)
Calculate the difference between the
expansions of H(zi) and the expectation of it,
square it and then take expectation
Comparison:
a.Calculate the difference between the asymptotic
variances:
b.To see which method is more accurate, the following
conditions can be checked for the i.i.d. case
(conditions for i.n.i.d. case is provided in the paper):
(i)
(ii)
(iii)
(vi) has the same sign with
[ ]2
var ( ) var ( ) ( )i i g Hg z H z or V V  − − 
2 2
2
4 (μ) (μ) (μ) 0i i i i i ig g H   ′ ′− ≥
   
22 21
(μ) (μ) (μ) (μ) 0
4
i i i i i i i ig g g H
    ′ ′′ ′′+ − ≥ ÷    
(μ) (μ) 0g g′′ ≥
( )2
[ (μ)] (μ) (μ)i i i i i ig g g′ ′′+
6 2 4
(μ)σ(μ)i i i i iE z E z   − − −   
Criteria:
• Check if all the four conditions are met
• When (i), (ii), (iii) and (vi) are all fulfilled, the 2nd
-
derivative approximation method outperforms the
product-of-derivative method regarding the accuracy
• These criteria are the sufficient conditions.
Case Study:
(1) N(μ, ) with θ= μ
(2) N(μ, ) with θ=
(3) Signal-plus noise problem: N(0, +q(i)) with θ= ,
where q(i)=0.1*(i-10*floor(i/10))
What’s happening
in practice?2
σ
2
σ 2
σ
2
σ2
σ
3.1188e+08 3.6371 3.1812e-08
1.9226e+08 2.0632 1.9380e-08
1.6222 1.7628 3.3915
2 2
μ1,σ0.1= = 2 2
μ5,σ10= =2
μ0,σ1= =
2
var ( )g z 
 
p-value 7.7846e-05 3.2146e-05 9.4938e-05
2 2
μ1,σ0.1= = 2
μ0,σ1= = 2 2
μ5,σ10= =
[ ]var ( )H z
[ ]2
var ( ) var ( )g z H z 
 
Selected Numerical Results:
Case 2: [ ]2 8 1
var ( ) var ( ) 3 (2σ) 0g z H z −  − = • > 
Case Study 3:
6.1060e+07 1.0349 3.5680e-08
2.5194e+07 0.6584 2.0023e-08
2.783 1.5718 1.7820
p-value 0.0269 0.0025 2.3863e-07
[ ]2
var ( ) var ( )i ig z H z 
 
[ ]var ( )iH z
2
var ( )ig z 
 
2 2
σ0.1= 2
σ1 = 2 2
σ10=
2 2
σ0.1= 2
σ1 = 2 2
σ10=
[ ] 1002 1 2 4
1
var ( ) var ( ) 3 (200) (σ) 0i i ii
g z H z q− −
=
  − = • + >  ∑
Conclusions:
• We decide which one of the product of derivative
approximation and the 2nd
-order derivative
approximation, is more accurate based on criteria (i),
(ii), (iii) and (vi) for the i.i.d. case
• Similar criteria for the i.n.i.d. scenario exists, which
can be found in the full paper
• For the normal density functions we considered in
our case study, we show that the Hessian methods
outperforms the gradient based method !
Thank you for listening !
References:
[1] Guo, S. (2014), "Comparison of Accuracy for Methods to
Approximate Fisher Information in the Scalar Case," M. S. project final
report, Department of Applied Mathematics and Statistics, Johns
Hopkins Univ.; available at arXiv: http://arxiv.org/abs/1501.00218.
[2] Guo, S. (2015), “Analysis of Data-Based Methods for
Approximating Fisher Information in the Scalar Case,” Proceedings of
the 49th Annual Conference on Information Sciences and
Systems, 18−20 March 2015, Baltimore, MD.
submitted to Proceedings of the 49th Annual Conference on Information Sciences and Systems, 18−20 March 2015, Baltimore, MD.
( )
2
22 2 2 2 2
4 4 2 2 3
2 4
var ( )
4 (μ)[ (μ)]σ[ (μ)] (μ) (μ) var (μ)
1
[ (μ)] var (μ) [ (μ)] [ (μ)] var (μ)
16
4 (μ)[ (μ)] (μ) (μ)
i
i
i i
i
g z
g g g g g z
g z g g z
g g g E z
 
 
 ′ ′ ′′≈ + + − 
   ′′ ′ ′′+ − + −   
 ′ ′′+ − 
( )( )2 2 6 2 41
[ (μ)] [ (μ)] (μ) (μ) (μ)σ(μ)
2
i ig g g g E z E z   ′′ ′ ′′+ + − − −   
[ ] [ ] [ ]2 22 21
var ( ) (μ)σ(μ) var (μ)
4
i iH z H H z ′ ′′≈ + − 
i.i.d. case:
2
μ( ),σvar( )i iE z z= =Define and let ( |θ) ( ), ( |θ) ( )g g H H• = • • = •
Consider the symmetric density family:
i.n.i.d. case:
( )
2
1
2 2
2 2 2 2
1 1
4 4 2 2 3
1 1
2
var ( )
4 (μ) (μ)σ[ (μ)] (μ) (μ) var (μ)
1
[ (μ)] var (μ) [ (μ)] [ (μ)] var (μ)
16
4 (μ)[ (μ)] (μ) (
n
i i
i
n n
i i i i i i i i i i i i i
i i
n n
i i i i i i i i i i
i i
i i i i i i
g z
g g g g g z
g z g g z
g g g E
=
= =
= =
 
 
 ′ ′ ′′  ≈ + + −  
′′ ′ ′′   + − + −   
′ ′′+
∑
∑ ∑
∑ ∑
( )( )
4
1
2 2 6 2 4
1
μ)
1
[ (μ)] [ (μ)] (μ) (μ) (μ)σ(μ)
2
n
i i
i
n
i i i i i i i i i i i i i
i
z
g g g g E z E z
=
=
 − 
′′ ′ ′′    + + − − −   
∑
∑
[ ]
2 2
2 2
1 1 1
1
var ( ) (μ)σ(μ) var (μ)
4
n n n
i i i i i i i i i
i i i
H z H H z
= = =
   ′ ′′  ≈ + −    ∑ ∑ ∑
Define and let
2
μ( ),σvar( )i i i iE z z= = ( |θ) ( ), ( |θ) ( )i i i ig g H H• = • • = •

Contenu connexe

Tendances

Green’s Function Solution of Non-homogenous Singular Sturm-Liouville Problem
Green’s Function Solution of Non-homogenous Singular Sturm-Liouville ProblemGreen’s Function Solution of Non-homogenous Singular Sturm-Liouville Problem
Green’s Function Solution of Non-homogenous Singular Sturm-Liouville ProblemIJSRED
 
Ejercicios varios de algebra widmar aguilar
Ejercicios varios de  algebra   widmar aguilarEjercicios varios de  algebra   widmar aguilar
Ejercicios varios de algebra widmar aguilarWidmar Aguilar Gonzalez
 
Paper id 71201914
Paper id 71201914Paper id 71201914
Paper id 71201914IJRAT
 
素数は孤独じゃない(番外編) 第13回 数学カフェ「素数!!」
素数は孤独じゃない(番外編) 第13回 数学カフェ「素数!!」素数は孤独じゃない(番外編) 第13回 数学カフェ「素数!!」
素数は孤独じゃない(番外編) 第13回 数学カフェ「素数!!」Junpei Tsuji
 
31350052 introductory-mathematical-analysis-textbook-solution-manual
31350052 introductory-mathematical-analysis-textbook-solution-manual31350052 introductory-mathematical-analysis-textbook-solution-manual
31350052 introductory-mathematical-analysis-textbook-solution-manualMahrukh Khalid
 
College algebra 7th edition by blitzer solution manual
College algebra 7th edition by blitzer solution manualCollege algebra 7th edition by blitzer solution manual
College algebra 7th edition by blitzer solution manualrochidavander
 
OPTIMIZATION OF DOPANT DIFFUSION AND ION IMPLANTATION TO INCREASE INTEGRATION...
OPTIMIZATION OF DOPANT DIFFUSION AND ION IMPLANTATION TO INCREASE INTEGRATION...OPTIMIZATION OF DOPANT DIFFUSION AND ION IMPLANTATION TO INCREASE INTEGRATION...
OPTIMIZATION OF DOPANT DIFFUSION AND ION IMPLANTATION TO INCREASE INTEGRATION...ijrap
 
On Optimization of Manufacturing of a Two-level Current-mode Logic Gates in a...
On Optimization of Manufacturing of a Two-level Current-mode Logic Gates in a...On Optimization of Manufacturing of a Two-level Current-mode Logic Gates in a...
On Optimization of Manufacturing of a Two-level Current-mode Logic Gates in a...BRNSS Publication Hub
 
A fixed point result in banach spaces
A fixed point result in banach spacesA fixed point result in banach spaces
A fixed point result in banach spacesAlexander Decker
 
Identification of unknown parameters and prediction of missing values. Compar...
Identification of unknown parameters and prediction of missing values. Compar...Identification of unknown parameters and prediction of missing values. Compar...
Identification of unknown parameters and prediction of missing values. Compar...Alexander Litvinenko
 
Hierarchical matrix techniques for maximum likelihood covariance estimation
Hierarchical matrix techniques for maximum likelihood covariance estimationHierarchical matrix techniques for maximum likelihood covariance estimation
Hierarchical matrix techniques for maximum likelihood covariance estimationAlexander Litvinenko
 
A common random fixed point theorem for rational inequality in hilbert space
A common random fixed point theorem for rational inequality in hilbert spaceA common random fixed point theorem for rational inequality in hilbert space
A common random fixed point theorem for rational inequality in hilbert spaceAlexander Decker
 
Re call basic operations in mathematics
Re call basic operations in mathematics Re call basic operations in mathematics
Re call basic operations in mathematics Nadeem Uddin
 
Low-rank matrix approximations in Python by Christian Thurau PyData 2014
Low-rank matrix approximations in Python by Christian Thurau PyData 2014Low-rank matrix approximations in Python by Christian Thurau PyData 2014
Low-rank matrix approximations in Python by Christian Thurau PyData 2014PyData
 
Numeros reales, inecuaciones y desigualdades
Numeros reales, inecuaciones y desigualdadesNumeros reales, inecuaciones y desigualdades
Numeros reales, inecuaciones y desigualdadesDanielaAngulo25
 

Tendances (17)

Green’s Function Solution of Non-homogenous Singular Sturm-Liouville Problem
Green’s Function Solution of Non-homogenous Singular Sturm-Liouville ProblemGreen’s Function Solution of Non-homogenous Singular Sturm-Liouville Problem
Green’s Function Solution of Non-homogenous Singular Sturm-Liouville Problem
 
Ejercicios varios de algebra widmar aguilar
Ejercicios varios de  algebra   widmar aguilarEjercicios varios de  algebra   widmar aguilar
Ejercicios varios de algebra widmar aguilar
 
Paper id 71201914
Paper id 71201914Paper id 71201914
Paper id 71201914
 
素数は孤独じゃない(番外編) 第13回 数学カフェ「素数!!」
素数は孤独じゃない(番外編) 第13回 数学カフェ「素数!!」素数は孤独じゃない(番外編) 第13回 数学カフェ「素数!!」
素数は孤独じゃない(番外編) 第13回 数学カフェ「素数!!」
 
31350052 introductory-mathematical-analysis-textbook-solution-manual
31350052 introductory-mathematical-analysis-textbook-solution-manual31350052 introductory-mathematical-analysis-textbook-solution-manual
31350052 introductory-mathematical-analysis-textbook-solution-manual
 
College algebra 7th edition by blitzer solution manual
College algebra 7th edition by blitzer solution manualCollege algebra 7th edition by blitzer solution manual
College algebra 7th edition by blitzer solution manual
 
Maths ms
Maths msMaths ms
Maths ms
 
OPTIMIZATION OF DOPANT DIFFUSION AND ION IMPLANTATION TO INCREASE INTEGRATION...
OPTIMIZATION OF DOPANT DIFFUSION AND ION IMPLANTATION TO INCREASE INTEGRATION...OPTIMIZATION OF DOPANT DIFFUSION AND ION IMPLANTATION TO INCREASE INTEGRATION...
OPTIMIZATION OF DOPANT DIFFUSION AND ION IMPLANTATION TO INCREASE INTEGRATION...
 
On Optimization of Manufacturing of a Two-level Current-mode Logic Gates in a...
On Optimization of Manufacturing of a Two-level Current-mode Logic Gates in a...On Optimization of Manufacturing of a Two-level Current-mode Logic Gates in a...
On Optimization of Manufacturing of a Two-level Current-mode Logic Gates in a...
 
A fixed point result in banach spaces
A fixed point result in banach spacesA fixed point result in banach spaces
A fixed point result in banach spaces
 
Identification of unknown parameters and prediction of missing values. Compar...
Identification of unknown parameters and prediction of missing values. Compar...Identification of unknown parameters and prediction of missing values. Compar...
Identification of unknown parameters and prediction of missing values. Compar...
 
Hierarchical matrix techniques for maximum likelihood covariance estimation
Hierarchical matrix techniques for maximum likelihood covariance estimationHierarchical matrix techniques for maximum likelihood covariance estimation
Hierarchical matrix techniques for maximum likelihood covariance estimation
 
A common random fixed point theorem for rational inequality in hilbert space
A common random fixed point theorem for rational inequality in hilbert spaceA common random fixed point theorem for rational inequality in hilbert space
A common random fixed point theorem for rational inequality in hilbert space
 
Re call basic operations in mathematics
Re call basic operations in mathematics Re call basic operations in mathematics
Re call basic operations in mathematics
 
Capitulo 5 Soluciones Purcell 9na Edicion
Capitulo 5 Soluciones Purcell 9na EdicionCapitulo 5 Soluciones Purcell 9na Edicion
Capitulo 5 Soluciones Purcell 9na Edicion
 
Low-rank matrix approximations in Python by Christian Thurau PyData 2014
Low-rank matrix approximations in Python by Christian Thurau PyData 2014Low-rank matrix approximations in Python by Christian Thurau PyData 2014
Low-rank matrix approximations in Python by Christian Thurau PyData 2014
 
Numeros reales, inecuaciones y desigualdades
Numeros reales, inecuaciones y desigualdadesNumeros reales, inecuaciones y desigualdades
Numeros reales, inecuaciones y desigualdades
 

En vedette

Certificate of Professional Continuity Practitioner
Certificate of Professional Continuity PractitionerCertificate of Professional Continuity Practitioner
Certificate of Professional Continuity PractitionerCharlie Shaw
 
23 raquel costa 24_rodolfo estima_vírus
23 raquel costa 24_rodolfo estima_vírus23 raquel costa 24_rodolfo estima_vírus
23 raquel costa 24_rodolfo estima_vírusidl8b24
 
TC_Electronic_M3000
TC_Electronic_M3000TC_Electronic_M3000
TC_Electronic_M3000Jon Gerber
 
Roland_V-Mixer
Roland_V-MixerRoland_V-Mixer
Roland_V-MixerJon Gerber
 
肥胖可能引致之疾病
肥胖可能引致之疾病肥胖可能引致之疾病
肥胖可能引致之疾病Perfectshape
 
Apresentacao BLV Monde 2015 Atualizada - PDF
Apresentacao BLV Monde 2015 Atualizada - PDFApresentacao BLV Monde 2015 Atualizada - PDF
Apresentacao BLV Monde 2015 Atualizada - PDFGlayson Santos
 
Sálvate de un naufragio Laboral: 35 años y ya eres viejo laboralmente
Sálvate de un naufragio Laboral: 35 años y ya eres viejo laboralmenteSálvate de un naufragio Laboral: 35 años y ya eres viejo laboralmente
Sálvate de un naufragio Laboral: 35 años y ya eres viejo laboralmenteCarlos Vargas H.
 
Labour Law and Employment in Ukraine - 2017 Guide
Labour Law and Employment in Ukraine - 2017 GuideLabour Law and Employment in Ukraine - 2017 Guide
Labour Law and Employment in Ukraine - 2017 GuideAccace
 
Zvýšenie minimálnej mzdy o 25 eur od 1.1.2016 | News Flash
Zvýšenie minimálnej mzdy o 25 eur od 1.1.2016 | News FlashZvýšenie minimálnej mzdy o 25 eur od 1.1.2016 | News Flash
Zvýšenie minimálnej mzdy o 25 eur od 1.1.2016 | News FlashAccace
 

En vedette (19)

Certificate of Professional Continuity Practitioner
Certificate of Professional Continuity PractitionerCertificate of Professional Continuity Practitioner
Certificate of Professional Continuity Practitioner
 
Baby
BabyBaby
Baby
 
23 raquel costa 24_rodolfo estima_vírus
23 raquel costa 24_rodolfo estima_vírus23 raquel costa 24_rodolfo estima_vírus
23 raquel costa 24_rodolfo estima_vírus
 
Test
TestTest
Test
 
TC_Electronic_M3000
TC_Electronic_M3000TC_Electronic_M3000
TC_Electronic_M3000
 
Biz Spark i co dalej
Biz Spark i co dalejBiz Spark i co dalej
Biz Spark i co dalej
 
Vida feliz1
Vida feliz1Vida feliz1
Vida feliz1
 
Iss
IssIss
Iss
 
Risco biologico
Risco biologicoRisco biologico
Risco biologico
 
Test
TestTest
Test
 
Roland_V-Mixer
Roland_V-MixerRoland_V-Mixer
Roland_V-Mixer
 
肥胖可能引致之疾病
肥胖可能引致之疾病肥胖可能引致之疾病
肥胖可能引致之疾病
 
BrandonSimmonsResume
BrandonSimmonsResumeBrandonSimmonsResume
BrandonSimmonsResume
 
Scott N. McWilliams Resume2
Scott N. McWilliams Resume2Scott N. McWilliams Resume2
Scott N. McWilliams Resume2
 
Apresentacao BLV Monde 2015 Atualizada - PDF
Apresentacao BLV Monde 2015 Atualizada - PDFApresentacao BLV Monde 2015 Atualizada - PDF
Apresentacao BLV Monde 2015 Atualizada - PDF
 
Sálvate de un naufragio Laboral: 35 años y ya eres viejo laboralmente
Sálvate de un naufragio Laboral: 35 años y ya eres viejo laboralmenteSálvate de un naufragio Laboral: 35 años y ya eres viejo laboralmente
Sálvate de un naufragio Laboral: 35 años y ya eres viejo laboralmente
 
Plastic roads
Plastic roadsPlastic roads
Plastic roads
 
Labour Law and Employment in Ukraine - 2017 Guide
Labour Law and Employment in Ukraine - 2017 GuideLabour Law and Employment in Ukraine - 2017 Guide
Labour Law and Employment in Ukraine - 2017 Guide
 
Zvýšenie minimálnej mzdy o 25 eur od 1.1.2016 | News Flash
Zvýšenie minimálnej mzdy o 25 eur od 1.1.2016 | News FlashZvýšenie minimálnej mzdy o 25 eur od 1.1.2016 | News Flash
Zvýšenie minimálnej mzdy o 25 eur od 1.1.2016 | News Flash
 

Similaire à ppt_tech

Smoothed Particle Galerkin Method Formulation.pdf
Smoothed Particle Galerkin Method Formulation.pdfSmoothed Particle Galerkin Method Formulation.pdf
Smoothed Particle Galerkin Method Formulation.pdfkeansheng
 
Measures of different reliability parameters for a complex redundant system u...
Measures of different reliability parameters for a complex redundant system u...Measures of different reliability parameters for a complex redundant system u...
Measures of different reliability parameters for a complex redundant system u...Alexander Decker
 
Development of a test statistic for testing equality of two means under unequ...
Development of a test statistic for testing equality of two means under unequ...Development of a test statistic for testing equality of two means under unequ...
Development of a test statistic for testing equality of two means under unequ...Alexander Decker
 
Attention Allocation to Partially Observable Heterogeneous Customers - with I...
Attention Allocation to Partially Observable Heterogeneous Customers - with I...Attention Allocation to Partially Observable Heterogeneous Customers - with I...
Attention Allocation to Partially Observable Heterogeneous Customers - with I...Avi Latner
 
Table 1
Table 1Table 1
Table 1butest
 
Boundary value problem and its application in i function of multivariable
Boundary value problem and its application in i function of multivariableBoundary value problem and its application in i function of multivariable
Boundary value problem and its application in i function of multivariableAlexander Decker
 
Maneuvering target track prediction model
Maneuvering target track prediction modelManeuvering target track prediction model
Maneuvering target track prediction modelIJCI JOURNAL
 
An investigation of inference of the generalized extreme value distribution b...
An investigation of inference of the generalized extreme value distribution b...An investigation of inference of the generalized extreme value distribution b...
An investigation of inference of the generalized extreme value distribution b...Alexander Decker
 
Stereographic Circular Normal Moment Distribution
Stereographic Circular Normal Moment DistributionStereographic Circular Normal Moment Distribution
Stereographic Circular Normal Moment Distributionmathsjournal
 
DISTANCE TWO LABELING FOR MULTI-STOREY GRAPHS
DISTANCE TWO LABELING FOR MULTI-STOREY GRAPHSDISTANCE TWO LABELING FOR MULTI-STOREY GRAPHS
DISTANCE TWO LABELING FOR MULTI-STOREY GRAPHSgraphhoc
 
ATT00001ATT00002ATT00003ATT00004ATT00005CARD.docx
ATT00001ATT00002ATT00003ATT00004ATT00005CARD.docxATT00001ATT00002ATT00003ATT00004ATT00005CARD.docx
ATT00001ATT00002ATT00003ATT00004ATT00005CARD.docxikirkton
 
On derivations of black scholes greek letters
On derivations of black scholes greek lettersOn derivations of black scholes greek letters
On derivations of black scholes greek lettersAlexander Decker
 
Statistics and Data Mining with Perl Data Language
Statistics and Data Mining with Perl Data LanguageStatistics and Data Mining with Perl Data Language
Statistics and Data Mining with Perl Data Languagemaggiexyz
 
Second-order Cosmological Perturbations Engendered by Point-like Masses
Second-order Cosmological Perturbations Engendered by Point-like MassesSecond-order Cosmological Perturbations Engendered by Point-like Masses
Second-order Cosmological Perturbations Engendered by Point-like MassesMaxim Eingorn
 

Similaire à ppt_tech (20)

Smoothed Particle Galerkin Method Formulation.pdf
Smoothed Particle Galerkin Method Formulation.pdfSmoothed Particle Galerkin Method Formulation.pdf
Smoothed Particle Galerkin Method Formulation.pdf
 
Measures of different reliability parameters for a complex redundant system u...
Measures of different reliability parameters for a complex redundant system u...Measures of different reliability parameters for a complex redundant system u...
Measures of different reliability parameters for a complex redundant system u...
 
Development of a test statistic for testing equality of two means under unequ...
Development of a test statistic for testing equality of two means under unequ...Development of a test statistic for testing equality of two means under unequ...
Development of a test statistic for testing equality of two means under unequ...
 
Attention Allocation to Partially Observable Heterogeneous Customers - with I...
Attention Allocation to Partially Observable Heterogeneous Customers - with I...Attention Allocation to Partially Observable Heterogeneous Customers - with I...
Attention Allocation to Partially Observable Heterogeneous Customers - with I...
 
Table 1
Table 1Table 1
Table 1
 
Boundary value problem and its application in i function of multivariable
Boundary value problem and its application in i function of multivariableBoundary value problem and its application in i function of multivariable
Boundary value problem and its application in i function of multivariable
 
Maneuvering target track prediction model
Maneuvering target track prediction modelManeuvering target track prediction model
Maneuvering target track prediction model
 
An investigation of inference of the generalized extreme value distribution b...
An investigation of inference of the generalized extreme value distribution b...An investigation of inference of the generalized extreme value distribution b...
An investigation of inference of the generalized extreme value distribution b...
 
Stereographic Circular Normal Moment Distribution
Stereographic Circular Normal Moment DistributionStereographic Circular Normal Moment Distribution
Stereographic Circular Normal Moment Distribution
 
QMC: Transition Workshop - Probability Models for Discretization Uncertainty ...
QMC: Transition Workshop - Probability Models for Discretization Uncertainty ...QMC: Transition Workshop - Probability Models for Discretization Uncertainty ...
QMC: Transition Workshop - Probability Models for Discretization Uncertainty ...
 
20320130406030
2032013040603020320130406030
20320130406030
 
Dielectric wave guide
Dielectric wave guideDielectric wave guide
Dielectric wave guide
 
DISTANCE TWO LABELING FOR MULTI-STOREY GRAPHS
DISTANCE TWO LABELING FOR MULTI-STOREY GRAPHSDISTANCE TWO LABELING FOR MULTI-STOREY GRAPHS
DISTANCE TWO LABELING FOR MULTI-STOREY GRAPHS
 
ATT00001ATT00002ATT00003ATT00004ATT00005CARD.docx
ATT00001ATT00002ATT00003ATT00004ATT00005CARD.docxATT00001ATT00002ATT00003ATT00004ATT00005CARD.docx
ATT00001ATT00002ATT00003ATT00004ATT00005CARD.docx
 
Au4201315330
Au4201315330Au4201315330
Au4201315330
 
On derivations of black scholes greek letters
On derivations of black scholes greek lettersOn derivations of black scholes greek letters
On derivations of black scholes greek letters
 
Sub1567
Sub1567Sub1567
Sub1567
 
QMC Program: Trends and Advances in Monte Carlo Sampling Algorithms Workshop,...
QMC Program: Trends and Advances in Monte Carlo Sampling Algorithms Workshop,...QMC Program: Trends and Advances in Monte Carlo Sampling Algorithms Workshop,...
QMC Program: Trends and Advances in Monte Carlo Sampling Algorithms Workshop,...
 
Statistics and Data Mining with Perl Data Language
Statistics and Data Mining with Perl Data LanguageStatistics and Data Mining with Perl Data Language
Statistics and Data Mining with Perl Data Language
 
Second-order Cosmological Perturbations Engendered by Point-like Masses
Second-order Cosmological Perturbations Engendered by Point-like MassesSecond-order Cosmological Perturbations Engendered by Point-like Masses
Second-order Cosmological Perturbations Engendered by Point-like Masses
 

ppt_tech

  • 1. Analysis of Data-Based Methods for Approximating Fisher Information in the Scalar Case Shenghan Guo Applied Math & Statistics, Johns Hopkins University
  • 3. Problem Setting: Random variables with density function p(z|θ), where the unknown parameter θ is scalar. Fisher information is the measurement of information amount carried by z regarding θ. 2 (θ) (θ| )nF E g =  z [ ](θ) (θ| ) ,nF E H=− z where log ( |θ) (θ| ) θ p g ∂ = ∂ z z and 2 2 log ( |θ) (θ| ) θ p H ∂ = ∂ z z 1 2[ , ,..., ]T nZ Z Z=z
  • 5. Variance Approximation for Maximum Likelihood Estimation (MLE) Asymptotic normality: For some common form of estimate , denote the true value of the parameter by . Then under modest conditions, ˆθn A classical use of Fisher Information θ* 1ˆ(θθ*) (0, (θ*)) dist nn N F− − →
  • 6. A Fact : It is difficult to get the true value of Fisher Information in practice because direct calculation could be difficult, or we don’t have the explicit forms of the gradient or the Hessian. Thus approximation is needed. Typical situations where approximation is needed: • State-space models with nuisance parameters (Spall, J. C. and Garner, J. P. 1990) • Non-i.i.d. projectile measurements and quantiles (Spall, J. C. and Maryak, J. L. 1992) • Inverse heat condition problems (Favennec, Y. 2007)
  • 7. • The unavailability of the true Fisher information in many practical cases makes a useful approximation method for this quantity important. • No prior studies have compared the approximation methods for Fisher information to decide which one is more accurate. • The convention of utilizing the certain approximation method (e.g. 2nd -derivative approximation to estimate Fisher information in the scalar case) needs theoretical support. Motivation:
  • 8. Two Methods to Approximate Fisher Information Number: I. Sums of Products of Derivatives: II. Negative Sum of 2nd Derivative: 2 1 1 (θ) (θ| ) n n i i i G g z n = = ∑ 1 1 (θ) (θ| ) n n i i i H H z n = = − ∑ The reality: Mostly people tend to use the 2nd derivative method instead of the product-of- derivatives method, which is more of a convention without valid theoretical support.
  • 9. My Way to Solve This Question:
  • 10. By the Central Limit Theorem: Let and , for independent and identically distributed random variables: ( ) ( )dist 2 (θ) (θ) 0,var ( )n in G F n N g z − →   ( ) [ ]( )dist (θ) (θ) 0,var ( )n in H F n N H z− → 2 2 ( ) (θ| )g g• = • ( ) (θ| )H H• = • nG nH We wonder which one of the estimates and is more accurate !
  • 11. For independent and non-identically distributed random variables: ( ) ( )dist (θ) (θ) 0,n gn G F n N V− → ( ) ( )dist (θ) (θ) 0,n Hn H F n N V−  → 1 2 1 lim var ( ) n g i iin V n g z− =→∞  =  ∑where and [ ]1 1 lim var ( ) n H i iin V n H z− =→∞ = ∑
  • 12. The Challenge: Calculation of the asymptotic variances in above CLT is infeasible because we cannot compute the variances analytically ! How to solve this problem? ----Estimation by Taylor expansion Assuming that our density function is at least twice- differentiable with respect to θ, we use the second- order Taylor expansion to approximate the asymptotic variances.
  • 13. 2 var ( )ig z    Take expectation of the Taylor expansion for the product of g(zi)
  • 14. [ ]var ( )iH z Take expectation of the Taylor expansion for H(zi) Calculate the difference between the expansions of H(zi) and the expectation of it, square it and then take expectation
  • 15. Comparison: a.Calculate the difference between the asymptotic variances: b.To see which method is more accurate, the following conditions can be checked for the i.i.d. case (conditions for i.n.i.d. case is provided in the paper): (i) (ii) (iii) (vi) has the same sign with [ ]2 var ( ) var ( ) ( )i i g Hg z H z or V V  − −  2 2 2 4 (μ) (μ) (μ) 0i i i i i ig g H   ′ ′− ≥     22 21 (μ) (μ) (μ) (μ) 0 4 i i i i i i i ig g g H     ′ ′′ ′′+ − ≥ ÷     (μ) (μ) 0g g′′ ≥ ( )2 [ (μ)] (μ) (μ)i i i i i ig g g′ ′′+ 6 2 4 (μ)σ(μ)i i i i iE z E z   − − −   
  • 16. Criteria: • Check if all the four conditions are met • When (i), (ii), (iii) and (vi) are all fulfilled, the 2nd - derivative approximation method outperforms the product-of-derivative method regarding the accuracy • These criteria are the sufficient conditions. Case Study: (1) N(μ, ) with θ= μ (2) N(μ, ) with θ= (3) Signal-plus noise problem: N(0, +q(i)) with θ= , where q(i)=0.1*(i-10*floor(i/10)) What’s happening in practice?2 σ 2 σ 2 σ 2 σ2 σ
  • 17. 3.1188e+08 3.6371 3.1812e-08 1.9226e+08 2.0632 1.9380e-08 1.6222 1.7628 3.3915 2 2 μ1,σ0.1= = 2 2 μ5,σ10= =2 μ0,σ1= = 2 var ( )g z    p-value 7.7846e-05 3.2146e-05 9.4938e-05 2 2 μ1,σ0.1= = 2 μ0,σ1= = 2 2 μ5,σ10= = [ ]var ( )H z [ ]2 var ( ) var ( )g z H z    Selected Numerical Results: Case 2: [ ]2 8 1 var ( ) var ( ) 3 (2σ) 0g z H z −  − = • > 
  • 18. Case Study 3: 6.1060e+07 1.0349 3.5680e-08 2.5194e+07 0.6584 2.0023e-08 2.783 1.5718 1.7820 p-value 0.0269 0.0025 2.3863e-07 [ ]2 var ( ) var ( )i ig z H z    [ ]var ( )iH z 2 var ( )ig z    2 2 σ0.1= 2 σ1 = 2 2 σ10= 2 2 σ0.1= 2 σ1 = 2 2 σ10= [ ] 1002 1 2 4 1 var ( ) var ( ) 3 (200) (σ) 0i i ii g z H z q− − =   − = • + >  ∑
  • 19. Conclusions: • We decide which one of the product of derivative approximation and the 2nd -order derivative approximation, is more accurate based on criteria (i), (ii), (iii) and (vi) for the i.i.d. case • Similar criteria for the i.n.i.d. scenario exists, which can be found in the full paper • For the normal density functions we considered in our case study, we show that the Hessian methods outperforms the gradient based method ! Thank you for listening !
  • 20. References: [1] Guo, S. (2014), "Comparison of Accuracy for Methods to Approximate Fisher Information in the Scalar Case," M. S. project final report, Department of Applied Mathematics and Statistics, Johns Hopkins Univ.; available at arXiv: http://arxiv.org/abs/1501.00218. [2] Guo, S. (2015), “Analysis of Data-Based Methods for Approximating Fisher Information in the Scalar Case,” Proceedings of the 49th Annual Conference on Information Sciences and Systems, 18−20 March 2015, Baltimore, MD. submitted to Proceedings of the 49th Annual Conference on Information Sciences and Systems, 18−20 March 2015, Baltimore, MD.
  • 21. ( ) 2 22 2 2 2 2 4 4 2 2 3 2 4 var ( ) 4 (μ)[ (μ)]σ[ (μ)] (μ) (μ) var (μ) 1 [ (μ)] var (μ) [ (μ)] [ (μ)] var (μ) 16 4 (μ)[ (μ)] (μ) (μ) i i i i i g z g g g g g z g z g g z g g g E z      ′ ′ ′′≈ + + −     ′′ ′ ′′+ − + −     ′ ′′+ −  ( )( )2 2 6 2 41 [ (μ)] [ (μ)] (μ) (μ) (μ)σ(μ) 2 i ig g g g E z E z   ′′ ′ ′′+ + − − −    [ ] [ ] [ ]2 22 21 var ( ) (μ)σ(μ) var (μ) 4 i iH z H H z ′ ′′≈ + −  i.i.d. case: 2 μ( ),σvar( )i iE z z= =Define and let ( |θ) ( ), ( |θ) ( )g g H H• = • • = • Consider the symmetric density family:
  • 22. i.n.i.d. case: ( ) 2 1 2 2 2 2 2 2 1 1 4 4 2 2 3 1 1 2 var ( ) 4 (μ) (μ)σ[ (μ)] (μ) (μ) var (μ) 1 [ (μ)] var (μ) [ (μ)] [ (μ)] var (μ) 16 4 (μ)[ (μ)] (μ) ( n i i i n n i i i i i i i i i i i i i i i n n i i i i i i i i i i i i i i i i i i g z g g g g g z g z g g z g g g E = = = = =      ′ ′ ′′  ≈ + + −   ′′ ′ ′′   + − + −    ′ ′′+ ∑ ∑ ∑ ∑ ∑ ( )( ) 4 1 2 2 6 2 4 1 μ) 1 [ (μ)] [ (μ)] (μ) (μ) (μ)σ(μ) 2 n i i i n i i i i i i i i i i i i i i z g g g g E z E z = =  −  ′′ ′ ′′    + + − − −    ∑ ∑ [ ] 2 2 2 2 1 1 1 1 var ( ) (μ)σ(μ) var (μ) 4 n n n i i i i i i i i i i i i H z H H z = = =    ′ ′′  ≈ + −    ∑ ∑ ∑ Define and let 2 μ( ),σvar( )i i i iE z z= = ( |θ) ( ), ( |θ) ( )i i i ig g H H• = • • = •