slides david kahle

Upload: yosue

Post on 03-Jun-2018

221 views

Category:

Documents


0 download

TRANSCRIPT

  • 8/12/2019 Slides David Kahle

    1/97

    Introducing Algebraic Statistics

    David Kahle

    Assistant Professor

    Department of Statistical Science

    June 3, 2013

    Introducing Algebraic StatisticsDavid Kahle

  • 8/12/2019 Slides David Kahle

    2/97

    Minimum Distance Estimation in Categorical CI Models2

    Outline

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    It is true that few unscientific people have this particular type of religious

    experience. Our poets do not write about it; our artists do not try to

    portray this remarkable thing. I don't know why. Is no one inspired by our

    present picture of the universe? This value of science remains unsung by

    singers: you are reduced to hearing not a song or poem, but an evening

    lecture about it. This is not yet a scientific age.

    Richard Feynman, The Value of Science

  • 8/12/2019 Slides David Kahle

    3/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    Whats algebraic statistics?

  • 8/12/2019 Slides David Kahle

    4/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    Whats algebraic statistics?

    Algebraic statistics is concerned with the development of techniques inalgebraic geometry, commutative algebra, and combinatorics, to addressproblems in statistics and its applications.

    Drton, Sturmfels, and Sullivant

    Lectures on Algebraic Statistics, 2009

  • 8/12/2019 Slides David Kahle

    5/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    Whats algebraic statistics?

    Algebraic statistics is concerned with the development of techniques inalgebraic geometry, commutative algebra, and combinatorics, to addressproblems in statistics and its applications.

    Drton, Sturmfels, and Sullivant

    Lectures on Algebraic Statistics, 2009

    Disclaimer :Im not an algebraist!

  • 8/12/2019 Slides David Kahle

    6/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    Whats algebraic statistics?

    Algebraic statistics is concerned with the development of techniques inalgebraic geometry, commutative algebra, and combinatorics, to addressproblems in statistics and its applications.

    Drton, Sturmfels, and Sullivant

    Lectures on Algebraic Statistics, 2009

    Disclaimer :Im not an algebraist!

    Great text :Cox, Little, and OSheas

    Ideals, Varieties, and Algorithms

  • 8/12/2019 Slides David Kahle

    7/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    Whats algebraic statistics?

    Algebraic statistics is concerned with the development of techniques inalgebraic geometry, commutative algebra, and combinatorics, to addressproblems in statistics and its applications.

    Drton, Sturmfels, and Sullivant

    Lectures on Algebraic Statistics, 2009

    Areas of application

    Multiway contingency tables

    Graphical models Factor analysis

    Structural equations models

    Statistical disclosure limitation

    Evolutionary biology

    Causal models Mixture models

    Bayesian integrals

    Disclaimer :Im not an algebraist!

  • 8/12/2019 Slides David Kahle

    8/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    My multivariatecalculus nightmare :

    5. Sketch x2 y2 z2= 1/2.

  • 8/12/2019 Slides David Kahle

    9/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    My multivariatecalculus nightmare :

    5. Sketch x2 y2 z2= 1/2.

    Lets ask !!

  • 8/12/2019 Slides David Kahle

    10/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    My multivariatecalculus nightmare :

    5. Sketch x2 y2 z2= 1/2.

    Lets ask !!

    ContourPlot3D[

    x^2 - y^2 - z^2 == 1/2,

    {x, -3, 3}, {y, -3, 3}, {z, -3, 3}

    ]

  • 8/12/2019 Slides David Kahle

    11/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    My multivariatecalculus nightmare :

    5. Sketch x2 y2 z2= 1/2.

  • 8/12/2019 Slides David Kahle

    12/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    My multivariatecalculus nightmare :

    5. Sketch x2 y2 z2= 1/2.

  • 8/12/2019 Slides David Kahle

    13/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    My multivariatecalculus nightmare :

    5. Sketch x2 y2 z2= 1/2.

    Over 600,000evaluations!

  • 8/12/2019 Slides David Kahle

    14/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    My multivariatecalculus nightmare :

    5. Sketch x2 y2 z2= 1/2.x2 y2 z2= 1/2

  • 8/12/2019 Slides David Kahle

    15/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    x2 y2 z2= 1/2

  • 8/12/2019 Slides David Kahle

    16/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    x2 y2 z2= 1/2

    Polynomial equation

  • 8/12/2019 Slides David Kahle

    17/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    x2 y2 z2= 1/2

    Polynomial equation

  • 8/12/2019 Slides David Kahle

    18/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    x2 y2 z2= 1/2

    Polynomial equation

  • 8/12/2019 Slides David Kahle

    19/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    x2 y2 z2= 1/2

    Polynomial equation

    This is called a realalgebraic set or anaffine variety.

    Th i i f f

  • 8/12/2019 Slides David Kahle

    20/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    The intersection of surfaces

    Where do these surfaces intersect?

    ATh i i f f

  • 8/12/2019 Slides David Kahle

    21/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    The intersection of surfaces

    Where do these surfaces intersect?

    BAYLORTh i i f f

  • 8/12/2019 Slides David Kahle

    22/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    The intersection of surfaces

    Where do these surfaces intersect?

    BAYLORTh i i f f

  • 8/12/2019 Slides David Kahle

    23/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    The intersection of surfaces

    Where do these surfaces intersect?

    Intersection = Solutions =

    BAYLORTh i t ti f f

  • 8/12/2019 Slides David Kahle

    24/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    The intersection of surfaces

    Geometrically....

    BAYLORTh i t ti f f

  • 8/12/2019 Slides David Kahle

    25/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    The intersection of surfaces

    BAYLORTh i t ti f f

  • 8/12/2019 Slides David Kahle

    26/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    The intersection of surfaces

    Secret : found with the cylindrical

    algebraic decomposition (CAD) algorithm

    BAYLORTh i t ti f f

  • 8/12/2019 Slides David Kahle

    27/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    The intersection of surfaces

    BAYLORThe intersection of surfaces

  • 8/12/2019 Slides David Kahle

    28/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    The intersection of surfaces

    BAYLORPolynomial ideals

  • 8/12/2019 Slides David Kahle

    29/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    Polynomial ideals

    Notation : the collection of polynomials

    with real coefficients is written

    BAYLORPolynomial ideals

  • 8/12/2019 Slides David Kahle

    30/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    Polynomial ideals

    Notation : the collection of polynomials

    with real coefficients is written

    Ring of polynomials

    BAYLORPolynomial ideals

  • 8/12/2019 Slides David Kahle

    31/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    Polynomial ideals

    Notation : the collection of polynomials

    with real coefficients is written

    Ring of polynomials

    Defn : The idealgenerated by the npolynomials

    is the set of polynomial combinations of thefks:

    BAYLORPolynomial ideals

  • 8/12/2019 Slides David Kahle

    32/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    Polynomial ideals

    Notation : the collection of polynomials

    with real coefficients is written

    Ring of polynomials

    Defn : The idealgenerated by the npolynomials

    is the set of polynomial combinations of thefks:

    BAYLORPolynomial ideals

  • 8/12/2019 Slides David Kahle

    33/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    Polynomial ideals

    Defn : The idealgenerated by the npolynomials

    is the set of polynomial combinations of thefks:

    BAYLORPolynomial ideals

  • 8/12/2019 Slides David Kahle

    34/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    Polynomial ideals

    Defn : The idealgenerated by the npolynomials

    is the set of polynomial combinations of thefks:

    Giant set of polynomials.

    BAYLORPolynomial ideals

  • 8/12/2019 Slides David Kahle

    35/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    Polynomial ideals

    Defn : The idealgenerated by the npolynomials

    is the set of polynomial combinations of thefks:

    Giant set of polynomials.

    Think : vector space.

    BAYLORPolynomial ideals

  • 8/12/2019 Slides David Kahle

    36/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    Polynomial ideals

    Defn : The idealgenerated by the npolynomials

    is the set of polynomial combinations of thefks:

    Giant set of polynomials.

    Think : vector space.

    Think : basis vectors.

    BAYLORPolynomial ideals

  • 8/12/2019 Slides David Kahle

    37/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    Polynomial ideals

    Defn : The idealgenerated by the npolynomials

    is the set of polynomial combinations of thefks:

    Giant set of polynomials.

    Think : vector space.

    Think : basis vectors.

    Fact : If are mpolynomials which generate

    the same ideal, then

    BAYLORPolynomial ideals

  • 8/12/2019 Slides David Kahle

    38/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    Polynomial ideals

    Defn : The idealgenerated by the npolynomials

    is the set of polynomial combinations of thefks:

    Giant set of polynomials.

    Think : vector space.

    Think : basis vectors.

    Fact : If are mpolynomials which generate

    the same ideal, then

    Think : a different basis.

    BAYLORPolynomial ideals

  • 8/12/2019 Slides David Kahle

    39/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    Polynomial ideals

    Defn : The idealgenerated by the npolynomials

    is the set of polynomial combinations of thefks:

    Giant set of polynomials.

    Think : vector space.

    Think : basis vectors.

    Fact : If are mpolynomials which generate

    the same ideal, then

    Think : a different basis.

    BAYLORPolynomial ideals

  • 8/12/2019 Slides David Kahle

    40/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    Polynomial ideals

    Fact : If are mpolynomials which generate

    the same ideal, then

    BAYLORPolynomial ideals

  • 8/12/2019 Slides David Kahle

    41/97

    u n i v e r s i t y

    BAYLOR

    Introducing Algebraic StatisticsDavid Kahle

    Polynomial ideals

    Fact : If are mpolynomials which generate

    the same ideal, then

    BAYLORPolynomial ideals

  • 8/12/2019 Slides David Kahle

    42/97

    u n i v e r s i t y

    Introducing Algebraic StatisticsDavid Kahle

    Polynomial ideals

    Fact : If are mpolynomials which generate

    the same ideal, then

    A Grbner basis is a good selection of thegks computed with Buchbergers algorithm

    or Faugeres F4 algorithm

    BAYLORPolynomial ideals

  • 8/12/2019 Slides David Kahle

    43/97

    u n i v e r s i t y

    Introducing Algebraic StatisticsDavid Kahle

    Polynomial ideals

    Fact : If are mpolynomials which generate

    the same ideal, then

    BAYLORPolynomial ideals

  • 8/12/2019 Slides David Kahle

    44/97

    u n i v e r s i t y

    Introducing Algebraic StatisticsDavid Kahle

    Polynomial ideals

    BAYLORPolynomial ideals

  • 8/12/2019 Slides David Kahle

    45/97

    u n i v e r s i t y

    Introducing Algebraic StatisticsDavid Kahle

    Polynomial ideals

    ... help find exactglobal solutions byelimination(but even in moderately sized problems can become intractable)

    Grbner bases can...

    ... identify if a polynomialfis in the ideal.

    ... allow for implicitization of a rationallyparameterized variety.

    ... compute related algebraic and geometric structures (e.g. the radical of the ideal)

    u n i v e r s i t y

    BAYLOR

  • 8/12/2019 Slides David Kahle

    46/97

    u n i v e r s i t y

    Introducing Algebraic StatisticsDavid Kahle

    Differentialgeometrystudies smooth manifoldsvia

    smooth topological structureAlgebraicgeometrystudies varietiesvia algebraic

    structure

    u n i v e r s i t y

    BAYLOR

  • 8/12/2019 Slides David Kahle

    47/97

    u n i v e r s i t y

    Introducing Algebraic StatisticsDavid Kahle

    Differentialgeometrystudies smooth manifoldsvia

    smooth topological structureAlgebraicgeometrystudies varietiesvia algebraic

    structure

    u n i v e r s i t y

    BAYLOR

  • 8/12/2019 Slides David Kahle

    48/97

    u n i v e r s i t y

    Introducing Algebraic StatisticsDavid Kahle

    Differentialgeometrystudies smooth manifoldsvia

    smooth topological structureAlgebraicgeometrystudies varietiesvia algebraic

    structure

    u n i v e r s i t y

    BAYLOR

  • 8/12/2019 Slides David Kahle

    49/97

    u n i v e r s i t y

    Introducing Algebraic StatisticsDavid Kahle

    Differentialgeometrystudies smooth manifoldsvia

    smooth topological structureAlgebraicgeometrystudies varietiesvia algebraic

    structure

    Real algebraic geometryis the study of real (semi)algebraic sets

    u n i v e r s i t y

    BAYLOR

  • 8/12/2019 Slides David Kahle

    50/97

    u n i v e r s i t y

    Introducing Algebraic StatisticsDavid Kahle

    Differentialgeometrystudies smooth manifoldsvia

    smooth topological structure

    Algebraicgeometrystudies varietiesvia algebraic

    structure

    A semialgebraic setis a finite union of sets of the form

    Real algebraic geometryis the study of real (semi)algebraic sets

    u n i v e r s i t y

    BAYLOR(Early?) Mantra of algebraic statistics

  • 8/12/2019 Slides David Kahle

    51/97

    u n i v e r s i t y

    Introducing Algebraic StatisticsDavid Kahle

    Fact 1 : Statistical inference depends on thegeometryof the parameter space(the model)

    ( y ) g

    E.g. regularity conditions, smoothness, etc.

    u n i v e r s i t y

    BAYLOR(Early?) Mantra of algebraic statistics

  • 8/12/2019 Slides David Kahle

    52/97

    Introducing Algebraic StatisticsDavid Kahle

    Fact 1 : Statistical inference depends on thegeometryof the parameter space(the model)

    Fact 2 : Ifthe parameter spaceis a semialgebraic set,

    then

    statistical inference can be analyzed with

    algebraic methods

    ( y ) g

    E.g. regularity conditions, smoothness, etc.

    u n i v e r s i t y

    BAYLOR(Early?) Mantra of algebraic statistics

  • 8/12/2019 Slides David Kahle

    53/97

    Introducing Algebraic StatisticsDavid Kahle

    Fact 1 : Statistical inference depends on thegeometryof the parameter space(the model)

    Fact 2 : Ifthe parameter spaceis a semialgebraic set,

    then

    statistical inference can be analyzed with

    algebraic methods

    ( y ) g

    E.g. regularity conditions, smoothness, etc.

    Statistical models are algebraic varieties

    u n i v e r s i t y

    BAYLOREx 1 : A parametric ASM

  • 8/12/2019 Slides David Kahle

    54/97

    Introducing Algebraic StatisticsDavid Kahle

    A simple experiment

    p

    u n i v e r s i t y

    BAYLOREx 1 : A parametric ASM

  • 8/12/2019 Slides David Kahle

    55/97

    Introducing Algebraic StatisticsDavid Kahle

    A simple experiment

    Requirements on the s

    p

    u n i v e r s i t y

    BAYLOREx 1 : A parametric ASM

  • 8/12/2019 Slides David Kahle

    56/97

    Introducing Algebraic StatisticsDavid Kahle

    A simple experiment

    1.

    Requirements on the s

    p

    u n i v e r s i t y

    BAYLOREx 1 : A parametric ASM

  • 8/12/2019 Slides David Kahle

    57/97

    Introducing Algebraic StatisticsDavid Kahle

    A simple experiment

    1.

    2.

    Requirements on the s

    u n i v e r s i t y

    BAYLOREx 1 : A parametric ASM

  • 8/12/2019 Slides David Kahle

    58/97

    Introducing Algebraic StatisticsDavid Kahle

    A simple experiment

    Requirements on the s

    1.

    2.

    u n i v e r s i t y

    BAYLOREx 1 : A parametric ASM

  • 8/12/2019 Slides David Kahle

    59/97

    Requirements on the s

    Introducing Algebraic StatisticsDavid Kahle

    A simple experiment

    1.

    2.

    This is a geometric structure!

    u n i v e r s i t y

    BAYLOREx 1 : A parametric ASM

  • 8/12/2019 Slides David Kahle

    60/97

    Requirements on the s

    Introducing Algebraic StatisticsDavid Kahle

    A simple experiment

    1.

    2.

    This is a geometric structure!

    u n i v e r s i t y

    BAYLOR

    Ex 1 : A parametric ASM

  • 8/12/2019 Slides David Kahle

    61/97

    Introducing Algebraic StatisticsDavid Kahle

    This is a geometric structure!

    Statistical models = subsets of the simplex

    u n i v e r s i t y

    BAYLOR

    Ex 1 : A parametric ASM

  • 8/12/2019 Slides David Kahle

    62/97

    Introducing Algebraic StatisticsDavid Kahle

    This is a geometric structure!

    Statistical models = subsets of the simplex

    Ex : The binomial distribution

    u n i v e r s i t y

    BAYLOR

    Ex 1 : A parametric ASM

  • 8/12/2019 Slides David Kahle

    63/97

    Introducing Algebraic StatisticsDavid Kahle

    This is a geometric structure!

    Statistical models = subsets of the simplex

    Ex : The binomial distribution

    u n i v e r s i t y

    BAYLOR

    Ex 1 : A parametric ASM

  • 8/12/2019 Slides David Kahle

    64/97

    Introducing Algebraic StatisticsDavid Kahle

    This is a geometric structure!

    Statistical models = subsets of the simplex

    Ex : The binomial distribution

    u n i v e r s i t y

    BAYLOR

    Ex 1 : A parametric ASM

  • 8/12/2019 Slides David Kahle

    65/97

    Introducing Algebraic StatisticsDavid Kahle

    This is a geometric structure!

    Statistical models = subsets of the simplex

    Observation :

    u n i v e r s i t y

    BAYLOR

    Ex 1 : A parametric ASM

  • 8/12/2019 Slides David Kahle

    66/97

    Introducing Algebraic StatisticsDavid Kahle

    This is a geometric structure!

    Statistical models = subsets of the simplex

    Observation :

    Implic

    itization

    u n i v e r s i t y

    BAYLOR

    Ex 1 : A parametric ASM

  • 8/12/2019 Slides David Kahle

    67/97

    Introducing Algebraic StatisticsDavid Kahle

    This is a geometric structure!

    Statistical models = subsets of the simplex

    Observation :

    Implic

    itization

    u n i v e r s i t y

    BAYLOR

    Ex 1 : A parametric ASM

  • 8/12/2019 Slides David Kahle

    68/97

    Introducing Algebraic StatisticsDavid Kahle

    This is a geometric structure!

    Statistical models = subsets of the simplex

    Observation :

    Implicitizat

    ion

    u n i v e r s i t y

    BAYLOR

    Ex 1 : A parametric ASM

  • 8/12/2019 Slides David Kahle

    69/97

    Introducing Algebraic StatisticsDavid Kahle

    This is a geometric structure!

    Statistical models = subsets of the simplex

    Observation :

    Implicitizat

    ion

    u n i v e r s i t y

    BAYLOREx 1 : A parametric ASM

  • 8/12/2019 Slides David Kahle

    70/97

    Introducing Algebraic StatisticsDavid Kahle

    The binomial model is the intersection of anaffine variety and the probability simplex!

    u n i v e r s i t y

    BAYLOREx 1 : A parametric ASM

  • 8/12/2019 Slides David Kahle

    71/97

    Introducing Algebraic StatisticsDavid Kahle

    Two descriptions of the binomial model

    Parametric description Implicit description

    Were used to this... ...but sometimes we get this.

    u n i v e r s i t y

    BAYLOREx 1 : A parametric ASM

  • 8/12/2019 Slides David Kahle

    72/97

    Introducing Algebraic StatisticsDavid Kahle

    Two descriptions of the binomial model

    Parametric description Implicit description

    Were used to this... ...but sometimes we get this.

    Parametric description Implicit description

    Possible (may be hard)

    Not always possible

    u n i v e r s i t y

    BAYLOREx 1 : A parametric ASM

  • 8/12/2019 Slides David Kahle

    73/97

    Introducing Algebraic StatisticsDavid Kahle

    A better visualization is available via

    barycentric coordinates

    3d to 2d

    A last note : Visualization

    ... but since were insidethe simplex, wecant see the varieties outside of it

    u n i v e r s i t y

    BAYLOREx 2 : An implicitly defined ASM

  • 8/12/2019 Slides David Kahle

    74/97

    Introducing Algebraic StatisticsDavid Kahle

    The 2 x 2 contingency table

    0 1

    0

    1

    X2

    X1

    where

    u n i v e r s i t y

    BAYLOREx 2 : An implicitly defined ASM

  • 8/12/2019 Slides David Kahle

    75/97

    Introducing Algebraic StatisticsDavid Kahle

    The 2 x 2 contingency table

    0 1

    0

    1

    X2

    X1

    where

    u n i v e r s i t y

    BAYLOREx 2 : An implicitly defined ASM

  • 8/12/2019 Slides David Kahle

    76/97

    Introducing Algebraic StatisticsDavid Kahle

    The 2 x 2 contingency table

    0 1

    0

    1

    X2

    X1

    where

    Requirements on the s

    1.

    2.

    u n i v e r s i t y

    BAYLOREx 2 : An implicitly defined ASM

  • 8/12/2019 Slides David Kahle

    77/97

    Introducing Algebraic StatisticsDavid Kahle

    The 2 x 2 contingency table

    0 1

    0

    1

    X2

    X1

    where

    Requirements on the s

    1.

    2. Points in/on the tetrahedronare distributions on the table!

    u n i v e r s i t y

    BAYLOREx 2 : An implicitly defined ASM

  • 8/12/2019 Slides David Kahle

    78/97

    Introducing Algebraic StatisticsDavid Kahle

    The 2 x 2 contingency table

    0 1

    0

    1

    X2

    X1

    where

    Requirements on the s

    1.

    2.

    2d

    3d

    Points in/on the tetrahedronare distributions on the table!

    u n i v e r s i t y

    BAYLOREx 2 : An implicitly defined ASM

  • 8/12/2019 Slides David Kahle

    79/97

    Introducing Algebraic StatisticsDavid Kahle

    0 1

    0

    1

    X2

    X1

    Ex : The independence model

    3d

    Statistical models = subsets of the simplex

    The 2 x 2 contingency table

    u n i v e r s i t y

    BAYLOREx 2 : An implicitly defined ASM

  • 8/12/2019 Slides David Kahle

    80/97

    Introducing Algebraic StatisticsDavid Kahle

    0 1

    0

    1

    X2

    X1

    Ex : The independence model

    3d

    Statistical models = subsets of the simplex

    means , where is the marginal probability of X1 = x1:

    The 2 x 2 contingency table

    u n i v e r s i t y

    BAYLOREx 2 : An implicitly defined ASM

  • 8/12/2019 Slides David Kahle

    81/97

    Introducing Algebraic StatisticsDavid Kahle

    0 1

    0

    1

    X2

    X1

    Ex : The independence model

    3d

    Statistical models = subsets of the simplex

    means , where is the marginal probability of X1 = x1:

    This holds for everycombination of x1and x2!

    The 2 x 2 contingency table

    u n i v e r s i t y

    BAYLOREx 2 : An implicitly defined ASM

  • 8/12/2019 Slides David Kahle

    82/97

    Introducing Algebraic StatisticsDavid Kahle

    0 1

    0

    1

    X2

    X1

    Ex : The independence model

    3d

    Statistical models = subsets of the simplex

    means , where is the marginal probability of X1 = x1:

    This holds for everycombination of x1and x2!

    The 2 x 2 contingency table

    u n i v e r s i t y

    BAYLOREx 2 : An implicitly defined ASM

  • 8/12/2019 Slides David Kahle

    83/97

    Introducing Algebraic StatisticsDavid Kahle

    0 1

    0

    1

    X2

    X1

    Ex : The independence model

    3d

    Statistical models = subsets of the simplex

    means , where is the marginal probability of X1 = x1:

    The 2 x 2 contingency table

    u n i v e r s i t y

    BAYLOREx 2 : An implicitly defined ASM

  • 8/12/2019 Slides David Kahle

    84/97

    Introducing Algebraic StatisticsDavid Kahle

    0 1

    0

    1

    X2

    X1

    Ex : The independence modelStatistical models = subsets of the simplex

    means , where is the marginal probability of X1 = x1:

    The 2 x 2 contingency table

    u n i v e r s i t y

    BAYLOREx 2 : An implicitly defined ASM

  • 8/12/2019 Slides David Kahle

    85/97

    Introducing Algebraic StatisticsDavid Kahle

    0 1

    0

    1

    X2

    X1

    Ex : The independence modelStatistical models = subsets of the simplex

    means , where is the marginal probability of X1 = x1:

    The 2 x 2 contingency table

    u n i v e r s i t y

    BAYLOREx 2 : An implicitly defined ASM

  • 8/12/2019 Slides David Kahle

    86/97

    Introducing Algebraic StatisticsDavid Kahle

    0 1

    0

    1

    X2

    X1

    Ex : The independence modelStatistical models = subsets of the simplex

    means , where is the marginal probability of X1 = x1:

    The 2 x 2 contingency table

    u n i v e r s i t y

    BAYLOREx 3 : A Gaussian ASM

  • 8/12/2019 Slides David Kahle

    87/97

    Introducing Algebraic StatisticsDavid Kahle

    Do the inferential methods change?

    For example, typically the MLE isasymptotically normal...

    u n i v e r s i t y

    BAYLOREx 3 : A Gaussian ASM

  • 8/12/2019 Slides David Kahle

    88/97

    Introducing Algebraic StatisticsDavid Kahle

    Simulation :1. Draw n= 50, 1000 samples from

    the bivariate normaldistribution with 0 mean.

    2. Compute MLE for mean (closestpoint to the parameter space)

    u n i v e r s i t y

    BAYLOREx 3 : A Gaussian ASM

  • 8/12/2019 Slides David Kahle

    89/97

    Introducing Algebraic StatisticsDavid Kahle

    One sample, its mean, and the MLE

    u n i v e r s i t y

    BAYLOREx 3 : A Gaussian ASM

  • 8/12/2019 Slides David Kahle

    90/97

    Introducing Algebraic StatisticsDavid Kahle

    Simulation :1. Draw n= 50, 1000 samples from

    the bivariate normaldistribution with 0 mean.

    2. Compute MLE for mean (closestpoint to the parameter space)

    Do N= 2000times

    u n i v e r s i t yBAYLOREx 3 : A Gaussian ASM

  • 8/12/2019 Slides David Kahle

    91/97

    Introducing Algebraic StatisticsDavid Kahle

    n= 50

    u n i v e r s i t yBAYLOREx 3 : A Gaussian ASM

  • 8/12/2019 Slides David Kahle

    92/97

    Introducing Algebraic StatisticsDavid Kahle

    n= 1000

    u n i v e r s i t yBAYLOREx 3 : A Gaussian ASM

  • 8/12/2019 Slides David Kahle

    93/97

    Introducing Algebraic StatisticsDavid Kahle

    Do the inferential methods change?

    For example, typically the MLE isasymptotically normal...

    Conclusion :

    Problems occur at singularities

    which dont ever go away.

    Problems happen near singularitieseven for relatively large n

    u n i v e r s i t yBAYLOR

  • 8/12/2019 Slides David Kahle

    94/97

    Introducing Algebraic StatisticsDavid Kahle

    Algebraic statistical models

    Defined in reference to another family of models (e.g. exponential family)

    Defined as semialgebraic sets on the parameter space.

    u n i v e r s i t yBAYLOR

  • 8/12/2019 Slides David Kahle

    95/97

    Introducing Algebraic StatisticsDavid Kahle

    Algebraic statistical models

    Defined in reference to another family of models (e.g. exponential family)

    Defined as semialgebraic sets on the parameter space.

    Drton and Sullivant, Statistica Sinica, 2007

    u n i v e r s i t yBAYLOR

  • 8/12/2019 Slides David Kahle

    96/97

    Introducing Algebraic StatisticsDavid Kahle

    Summary : Algebraic statistical models are an interestingand general class of models, but require care with many of theclassical aspects of inference...

    u n i v e r s i t yBAYLOR

  • 8/12/2019 Slides David Kahle

    97/97

    Summary : Algebraic statistical models are an interestingand general class of models, but require care with many of theclassical aspects of inference...

    ... with cool pictures.

    Thank you!