Open Problems in Algebraic Statistics

Transcription

1 Open Problems inalgebraic Statistics p. Open Problems in Algebraic Statistics BERND STURMFELS UNIVERSITY OF CALIFORNIA, BERKELEY and TECHNISCHE UNIVERSITÄT BERLIN Advertisement Oberwolfach Seminar Algebraic Statistics with Mathias Drton and Seth Sullivant May 11-17, 2008

2 Open Problems inalgebraic Statistics p. Models with Hidden Variables A Concrete Problem: Consider the variety of tables of tensor rank at most 4. Do the known polynomial invariants of degree five and nine suffice to cut out this variety? [ASCB, Conjecture 3.24, page 102] The General Problem: Study the geometry and commutative algebra of graphical models with hidden random variables. Construct these varieties by gluing familiar secant varieties, and by applying representation theory.

3 Open Problems inalgebraic Statistics p. My favorite statistical model lives in the 64-dimensional space C 4 C 4 C 4 of tables (p ijk ), where i,j,k {A,C,G,T}. It is the variety with parametric representation p ijk = ρ Ai σ Aj θ Ak + ρ Ci σ Cj θ Ck + ρ Gi σ Gj θ Gk + ρ Ti σ Tj θ Tk Our problem is to compute its homogeneous prime ideal I in the polynomial ring with 64 unknowns, Q [ p AAA,p AAC,p AAT,...,p TTG,p TTT ].

4 Open Problems inalgebraic Statistics p. Why (Pure) Mathematics? Mathematics is the Art of Giving the Same Name to Different Things (H. Poincaré). the set of tables of tensor rank 4 mixture of independent random variables the naive Bayes model the CI model [X 1 X 2 X 3 Y ]. third secant variety of Segre variety P 3 P 3 P 3 general Markov model for phylogenetic tree K 1,3 entanglement of pure states (mean fields)

5 Open Problems inalgebraic Statistics p. Invariants of Degree Five Consider any subtable of (p ijk ) and let A,B,C be the 4 4-slices gotten by fixing i. On our model, the following identity of 4 4-matrices holds: A B 1 C = C B 1 A After clearing the denominator det(b), this gives 16 quintic polynomials which lie in our prime ideal I. Proposition 1. The space of quintic polynomials in I has dimension As a GL(C 4 ) 3 -module, it equals S 311 (C 4 ) S 2111 (C 4 ) S 2111 (C 4 ) S 2111 (C 4 ) S 311 (C 4 ) S 2111 (C 4 ) S 2111 (C 4 ) S 2111 (C 4 ) S 311 (C 4 )

6 Open Problems inalgebraic Statistics p. Invariants of Degree Nine Consider any subtable (A,B,C) of (p ijk ). On our model, the following 3 3-matrix is singular: A B 1 C C B 1 A The numerator of its determinant is a degree nine polynomial in I known as the Strassen invariant. Proposition 2. The GL(C 4 ) 3 -submodule of I 9 generated by the Strassen invariant is not contained in I 5. This module has dimension 8000 and it equals S 333 (C 4 ) S 333 (C 4 ) S 333 (C 4 ). References: [Garcia-Stillman-St], [Allman-Rhodes], [Landsberg-Manivel], [Landsberg-Weyman]

7 Open Problems inalgebraic Statistics p. Maximum Likelihood A Concrete Problem: Characterize all projective varieties whose maximum likelihood degree is equal to one. [HKS: Solving the Likelihood Equations, Problem 13] The General Problem: Study the geometry of maximum likelihood estimation for algebraic statistical models. What makes a model nice? Is the ML degree related to convergence properties of the EM algorithm?

8 Open Problems inalgebraic Statistics p. MLE in Projective Space Fix projective space P n with coordinates (p 0 : p 1 : : p n ). The n-dimensional probability simplex is identified with P n 0. For any data vector (u 0,u 1,...,u n ) N n+1 we consider the likelihood function L = p 0 u 0 p 1 u 1 p 2 u 2 p n u n (p 0 +p 1 + +p n ) u 0+u 1 + +u n This is a well-defined rational function on P n. Q: What are the critical points of this function?

9 Open Problems inalgebraic Statistics p. Algebraic Statistical Models are subvarieties of the projective space P n. The maximum likelihood degree of a model M is the number of critical points of the restriction to M of L = p 0 u 0 p 1 u 1 p 2 u 2 p n u n (p 0 +p 1 + +p n ) u 0+u 1 + +u n Here we only count critical points that are not poles or zeros, and u 0,u 1,...,u n are assumed to be generic. If M is smooth and the divisor of L is normal crossing then there is a formula for the ML degree. [Catanese-Hoşten-Khetan-St.] Otherwise, use resolution of singularities????

10 Open Problems inalgebraic Statistics p. 1 Determinantal Varieties are models M that are specified by imposing rank conditions on a matrix of unknowns. ( ) p00 p Example. 2 2-matrices 01 of rank 1. p 10 p 11 Independence for two binary random variables. The ML degree is one, i.e., the MLE is a rational function in the data (Take normalized row and column sums) Example. 3 3-matrices (p ij ) of rank 2. Mixture of two ternary random variables. The ML degree is ten. Open Problem. Determine the ML degree of the variety of m n-matrices of rank r.

11 Open Problems inalgebraic Statistics p. 1 The 100 Swiss Francs Problem Our data are two aligned DNA sequences... ATCACCAAACATTGGGATGCCTGTGCATTTGCAAGCGGCT ATGAGTCTTAAACGCTGGCCATGTGCCATCTTAGACAGCG... test the hypothesis that these two sequences were generated by DiaNA using one biased coin and four tetrahedral dice... [ASCB, Example 1.16] Model: Positive 4 4-matrices (p ij ) of rank True or False: The matrix (ˆp ) ij = maximizes L = ( i p ii ) 4 ( i j p ij ) 2 ( i,j p ij ) 40.

12 Open Problems inalgebraic Statistics p. 1 Gaussian CI Models A Concrete Problem: Which sets of almost-principal minors can be zero for a positive definite symmetric 5 5-matrix? [Matùš, Studený, Drton, Sullivant, Parrilo,...] The General Problem: Study the geometry of conditional independence models for multivariate Gaussian random variables. Which semigraphoids can be realized by Gaussians?

13 Open Problems inalgebraic Statistics p. 1 Almost-Principal Minors A multivariate Gaussian with mean zero is specified by its covariance matrix Σ = (σ ij ). This is a positive definite n n-matrix: all principal minors are positive. An almost-principal minor of Σ has row indices {i} K and column indices {j} K for some K [n] and i,j [n]\k. It is denoted [i j K ]. Lemma. The determinant [i j K ] is zero if and only if the random variable i is independent of the random variable j given the joint random variable K.

14 Open Problems inalgebraic Statistics p. 1 Subvarieties of the PSD cone Let PSD n denote the ( ) n+1 2 -dimensional cone of positive semidefinite symmetric n n-matrices. A Gaussian conditional independence model is a subvariety of PSD n defined by equations [i j K ] = 0. In algebraic geometry, we would study these varieties over the complex numbers. Here we are interested in their real locus, and how it intersects the cone PSD n.

15 Open Problems inalgebraic Statistics p. 1 A Cyclic Example Let n = 5 and consider the CI model given by [ 1 2 {3} ] = σ 12 σ 33 σ 13 σ 23 [ 2 3 {4} ] = σ 23 σ 44 σ 24 σ 34 [ 3 4 {5} ] = σ 34 σ 55 σ 35 σ 45 [ 4 5 {1} ] = σ 45 σ 11 σ 14 σ 15 [ 5 1 {2} ] = σ 15 σ 22 σ 25 σ 12 This variety has two irreducible components σ 12,σ 23,σ 34,σ 45,σ 15 a toric variety of degree 31, which intersects PSD 5 only in the set of rank one matrices. Corollary: For Gaussians, the five given statements imply [ 1 2 ], [ 2 3 ], [ 3 4 ], [ 4 5 ], [ 5 1 ].

16 Open Problems inalgebraic Statistics p. 1 The Entropy Map The submodular cone SubMod n is solution set of the following system of linear inequalities in R 2n : H {i} K + H {j} K H {i,j} K + H K Taking the logarithms of all 2 n principal minors of any positive definite matrix defines the entropy map H : SDP n SubMod n Proposition. The Gaussian CI models are the inverse images of the faces of the polyhedral cone SubMod n. Problem: Characterize the image of H.

17 Open Problems inalgebraic Statistics p. 1 Conclusion This talk presented three concrete open problems: 1. Consider the variety of tables of tensor rank at most 4. Do the known polynomial invariants of degree five and nine suffice to cut out this variety? 2. Characterize all projective varieties whose maximum likelihood degree is equal to one. 3. Which sets of almost-principal minors can be zero for a positive definite symmetric 5 5-matrix?

18 Open Problems inalgebraic Statistics p. 1 Bonus Problem: Rational Points Consider n discrete random variables X 1,X 2,...,X n with d 1,d 2,...,d n states. Any collection of CI statements defines a determinantal variety in the space of tables, C d 1 C d 2 C d n. The corresponding strict CI variety is the set of tables for which the given CI statements hold but all other CI statement do not hold. Problem: Does every strict CI variety have a Q-rational point? What if d 1,d 2,...,d n 0? [Matúš: Final Conclusions, CPC 8 (1999) p. 275]