Subset Selection. Ilse Ipsen. North Carolina State University, USA

Transcription

1 Subset Selection Ilse Ipsen North Carolina State University, USA

2 Subset Selection Given: real or complex matrix A integer k Determine permutation matrix P so that AP = ( A 1 }{{} k A 2 ) Important columns A 1 Columns of A 1 are very linearly independent Wannabe basis vectors Redundant columns A 2 Columns of A 2 are well represented by A 1

3 Outline Two applications Mathematical formulation Bounds Algorithms (two norm) Redundant columns (Frobenius norm) Randomized algorithms

4 First application Joint work with Tim Kelley Solution of nonlinear least squares problems Levenberg-Marquardt trust-region algorithm Difficulties: illconditioned or rank deficient Jacobians errors in residuals and Jacobians Improve conditioning with subset selection Damped driven harmonic oscillators allow us to construct different scenarios

5 Harmonic Oscillators m x + c x + k x = 0 displacement x(t) mass m damping constant c spring stiffness k Given: displacement measurements x j at time t j Want: parameters p = (m, c, k) Nonlinear least squares problem min x(t j, p) x j 2 p j

6 Nonlinear Least Squares Problem Residual R(p) = ( x(t 1, p) x 1 x(t 2, p) x 2... ) T Nonlinear Least Squares Problem min p f(p) where f(p) = 1 2 R(p)T R(p) At a minimizer p : f(p ) = 0 Gradient: f(p) = R (p) T R(p) Jacobian: R (p) Solve f(p) = 0 by Levenberg-Marquardt algorithm p new = p ( νi + R (p) T R (p)) 1 R (p) T R(p)

7 Jacobian Levenberg-Marquardt algorithm p new = p ( νi + R (p) T R (p)) 1 R (p) T R(p) But: Jacobian R (p) does not have full column rank m x + c x + k x = 0 for infinitely many p = (m, c, k) x + c m x + k m x = 0 m c x + x + k c x = 0 m k x + c k x + x = 0 Which parameter to keep fixed? Which column in the Jacobian is redundant? Subset Selection!

8 Second Application Scott Pope s Ph.D. thesis (NCState, 2009) Cardiovascular and respiratory modeling Identify parameters that are important for predicting blood flow and pressure Solve nonlinear least squares problem combined with subset selection Parameters identified as important: total systemic resistance cerebrovascular resistance arterial compliance time of peak systolic ventricular pressure

9 Mathematical Formulation of Subset Selection Given: real or complex matrix A with n columns integer k Determine permutation matrix P so that AP = ( A 1 }{{} k A }{{} 2 ) n k Important columns A 1 Columns of A 1 are very linearly independent Smallest singular value of A 1 is large Redundant columns A 2 Columns of A 2 are well represented by A 1 min Z A 1 Z A 2 is small (two norm)

10 Singular Value Decomposition (SVD) m n matrix A, m n AP = U Σ V Singular vectors: U T U = I m V T V = VV T = I n Singular values: Σ = σ 1... σ 1... σ n 0 σ n

11 Ideal Matrices for Subset Selection Singular Value Decomposition AP = ( A 1 }{{} k ( Σ1 A }{{} 2 ) = U n k Σ 2 ) V If exists permutation P so that V = I then Important columns large singular values of A ( ) Σ1 A 1 = U and σ 0 i (A 1 ) = σ i (A) 1 i k Redundant columns small singular values of A ( ) 0 A 2 = U and min A 1 Z A 2 = A 2 = σ k+1 (A) Z Σ 2

12 Subset Selection Requirements Important columns A 1 k columns of A 1 should be very linearly independent Smallest singular value σ k (A 1 ) should be large for some γ σ k (A)/γ σ k (A 1 ) σ k (A) Redundant columns A 2 Columns of A 2 should be well represented by A 1 min Z A 1 Z A 2 should be small (two norm) for some γ σ k+1 (A) min Z A 1 Z A 2 γ σ k+1 (A)

13 Bounds for Subset Selection Singular value decomposition AP = ( A 1 }{{} k A }{{} 2 ) = U n k ( Σ1 Σ 2 ) ( ) V11 V 12 V 21 V 22 Important columns A 1 σ i (A) V 1 11 σ i(a 1 ) σ i (A) for all 1 i k Redundant columns A 2 σ k+1 (A) min Z A 1 Z A 2 V 1 11 σ k+1(a)

14 Matrix ( V 11 How Small Can V 1 11 Be? V 12 ) is k n with orthonormal rows If V 11 is nonsingular then V V 1 11 V 12 2 Follows from I = V 11 V T 11 + V 12V T 12 If det(v 11 ) is maximal then V 1 11 V 12 k(n k) Follows from Cramer s rule: (V 1 11 V 12) ij 1 There exists a permutation such that V k(n k)

15 Bounds for Subset Selection AP = ( A 1 }{{} k ( Σ1 A }{{} 2 ) = U n k Σ 2 ) ( ) V11 V 12 V 21 V 22 Permute right singular vectors so that det(v 11 ) maximal Important columns A 1 σ i (A) 1 + k(n k) σ i (A 1 ) σ i (A) for all 1 i k Redundant columns A 2 σ k+1 (A) min Z A 1 Z A k(n k) σ k+1 (A)

16 Algorithms for Subset Selection QR decomposition with column pivoting ( ) R11 R AP = Q 12 where Q T Q = I 0 R 22 Important columns ( ) R11 Q = A 0 1 and σ i (A 1 ) = σ i (R 11 ) 1 i k Redundant columns ( ) R12 Q = A 2 min A 1 Z A 2 = R 22 Z R 22

17 QR Decomposition with Column Pivoting Determines permutation matrix P so that ( ) R11 R AP = Q 12 0 R 22 where R 11 well-conditioned R 22 small Numerical rank of A is k QR decomposition reveals rank

18 Rank Revealing QR (RRQR) Decompositions Businger & Golub (1965) QR with column pivoting Faddev, Kublanovskaya & Faddeeva (1968) Golub, Klema & Stewart (1976) Gragg & Stewart (1976) Stewart (1984) Foster (1986) T. Chan (1987) Hong & Pan (1992) Chandrasekaran & Ipsen (1994) Gu & Eisenstat (1996) Strong RRQR

19 Input: Strong RRQR (Gu & Eisenstat 1996) m n matrix A, m n, integer k ( ) R11 R Output: AP = Q 12 0 R 22 R 11 is well conditioned R 11 is k k σ i (A) 1 + k(n k) σ i (R 11 ) σ i (A) 1 i k R 22 is small σ k+j (A) σ j (R 22 ) 1 + k(n k) σ k+j (A) Offdiagonal block not too large ( ) R 1 11 R 12 1 ij

20 Strong RRQR Algorithm 1 Compute some QR decomposition with column pivoting ( ) R11 R AP initial = Q 12 0 R 22 2 Repeat «R11 Exchange a column of with a column of 0 Update permutations P, retriangularize until det(r 11 ) stops increasing 3 Output: AP final = ( A 1 }{{} k A }{{} 2 ) n k R12 R 22 «Operation count: O ( mn 2) until det(r 11) stops increasing by n

21 Another Strong RRQR Algorithm In the spirit of Golub, Klema and Stewart (1976) 1 Compute SVD ( Σ1 A = U Σ 2 ) ( ) V1 V 2 2 Apply strong RRQR to V 1 : V 1 P = (V }{{} 11 k V }{{} 12 ) n k k(n k) σ i (V 11 ) 1 1 i k 3 Output: AP = ( A }{{} 1 k A }{{} 2 ) n k

22 Summary: Deterministic Subset Selection m n matrix A, m n, AP = ( A 1 }{{} k Important columns A 1 Redundant columns A 2 A }{{} 2 ) n k σ k (A)/p(k, n) σ k (A 1 ) σ k (A) σ k+1 (A) min Z A 1 Z A 2 p(k, n) σ k+1 (A) p(k, n) Depends on leading k right singular vectors Best known value: p(k, n) = 1 + k(n k) Algorithms: Rank revealing QR decompositions, SVD Operation count: O ( mn 2) for p(k, n) = p 1 + nk(n k)

23 AP = ( A 1 }{{} k RRQR: A }{{} 2 ) n k Redundant Columns min Z A 1 Z A 2 p(k, n) σ k+1 (A) Orthogonal projector onto range(a 1 ): A 1 A 1 min z A 1 Z A 2 = (I A 1 A 1 ) A 2 Largest among all small singular values σ k+1 (A) = Σ 2 RRQR: (I A 1 A 1 ) A 2 p(k, n) Σ 2

24 Subset Selection for Redundant Columns AP = ( A 1 }{{} k A }{{} 2 ) n k Among all ( n k) choices find A1 that minimizes (I A 1 A 1 ) A 2 ξ Best bounds: 2 norm (I A 1 A 1 ) A k(n k) Σ 2 2 Frobenius norm (I A 1 A 1 ) A 2 F k + 1 Σ 2 F

25 Frobenius Norm There exist k columns A 1 so that (I A 1 A 1 ) A 2 2 F (k + 1) σ j (A) 2 j k+1 Idea: Volume sampling Deshpande, Rademacher, Vempala & Wang 2006 (I A 1 A 1 ) A 2 2 F = j k+1 (I A 1A 1 ) a j 2 2 Volume Vol(a j ) = a j 2 Vol ( A 1 a j ) = Vol(A1 ) (I A 1 A 1 )a j 2 Volume and singular values i 1 <...<i k Vol (A i1...i k ) 2 = i 1 <...<i k σ i1 (A) 2... σ ik (A) 2

26 Maximizing Volumes Is Really Hard Given: matrix A with n columns of unit norm integer k real number ν [0, 1] Finding k columns A 1 of A such that Vol(A 1 ) ν is NP-hard There is no polynomial time approximation scheme [Civril & Magdon-Ismail, 2007]

27 Randomized Subset Selection Frieze, Kannan & Vempala 2004 Deshpande, Rademacher, Vempala & Wang 2006 Liberty, Woolfe, Martinsson, Rokhlin & Tygert 2007 Drineas, Mahoney & Muthukrishnan 2006, 2008 Boutsidis, Mahoney & Drineas 2009 Civril & Magdon-Ismail 2009 Applications Statistical data analysis: feature selection principal component analysis Pass efficient algorithms for large data sets

28 2-Phase Randomized Algorithm Boutsidis, Mahoney & Drineas Randomized Phase: Sample small number ( k log k) of columns 2 Deterministic Phase: Apply rank revealing QR to sampled columns With 70% probability: Two norm min Z A 1 Z A 2 2 O Frobenius norm min Z A 1 Z A 2 F O ( k 3/4 log 1/2 k (n k) 1/4) Σ 2 2 ( ) k log 1/2 k Σ 2 F

29 2-Phase Randomized Algorithm 1 Compute SVD ( Σ1 A = U ) ( ) V1 Σ 2 V 2 2 Randomized phase: Scale: V 1 V 1 D Sample c columns: 3 Deterministic phase: Apply RRQR to V s : 4 Output: AP s P d = ( A }{{} 1 k (V 1 D) P s = ( V s }{{} ĉ V s P d = ( V d }{{} k A }{{} 2 ) n k ) )

30 Randomized Phase (Frobenius Norm) Sampling Column i of V 1 sampled with probability p i Probabilities p i = c ( ) (V1 ) i i n V 1 F Scaling Scaling matrix D = diag ( 1/ p / p n ) Scaled matrix V 1 D All columns have the same norm Columns sampled with probability 1/n Purpose of scaling makes sampling uniform makes expected values easier to compute

31 Analysis of 2-Phase Algorithm ( ) V1s D 1 Sample c columns (VD) P s = s V 2s D s 2 RRQR selects k columns (V 1s D s ) P d = ( V d ) Perturbation theory RRQR min Z A 1 Z A 2 F Σ 2 F + Σ 2 V 2s D s F σ k (V d ) With high probability σ k (V 1s D s ) 1/2 c k log k σ k (V d ) σ k(v 1s D s ) 1 + k(ĉ k) min Z A 1 Z A 2 F O Σ 2 V 2s D s F 4 Σ 2 F ( ) k log 1/2 k Σ 2 F

32 Expected Values of Frobenius Norms If D ii = 1/ p i then ( ) E X D 2 F = X 2 F Frobenius norm X D 2 F = trace(x D2 X T ) Linearity Scaling [ ] E X D 2 F = trace(x E [D 2] }{{} I X T ) [ ] E D 2 ii = p i 1 + (1 p i ) 0 = 1 p i

33 From Expected Values to Probability ( ) E X D 2 F = X 2 F Markov s inequality x = X D 2 F, Prob (x a) E(x)/a a = 10 E(x) With probability at most 1/10 X D 2 F 10 X 2 F With probability at least 9/10 X D 2 F 10 X 2 F

34 Issues with Randomized Algorithms How to choose c: 10 3 k log k, k log k, 17 k log k,...? We don t know the number of sampled columns ĉ Number of sampled columns can be too small: ĉ < k No information about singular values of important columns How often does one have to run the algorithm to get a good result? How accurately do the singular vectors and singular values have to be computed? How sensitive is the algorithm to the choice of probabilities? How does the randomized algorithm compare to the deterministic algorithms: accuracy, run time?

35 Summary Given: real or complex matrix A, integer k Want: AP = ( A }{{} 1 A 2 ) }{{} k n k Important columns A 1 Singular values close to k largest singular values of A Redundant columns A 2 Proj. of A 2 on range(a 1) 2,F smallest singular values of A Bounds depend on dominant k right singular vectors Deterministic algorithms: RRQR, SVD Randomized algorithm: 2 phases: 1. randomized sampling, 2. RRQR on samples Exact subset selection is hard