The number of singular vector tuples and approximation of symmetric tensors / 13

Transcription

1 The number of singular vector tuples and approximation of symmetric tensors Shmuel Friedland Univ. Illinois at Chicago Colloquium NYU Courant May 12, 2014 Joint results with Giorgio Ottaviani and Margaret Stawiska

2 Notations Indices: m = (m 1,..., m d ) N d, [m] := {1,..., m}, F = C, R, J = {j 1,..., j k } [d] Tensors: d i=1 Fm i = F m 1... m d = F m Contraction of T = [t i1,...,i d ] F m with X = [x ij1,...,i jk ] jp JF m jp : T X = i j p [m jp ],jp J t i 1,...,i d x ij1,...,i jk l [d]\j F m l Example T (x 1... x k 1 x k+1... x d ) = i j [m j ],j [d]\{k} t i 1,...,i d j [d]\{k} x i j,j is a vector in F m k T = T T - Hilbert-Schmidt norm of T R m

3 Singular values and vectors for tensors Introduced by Lek-Heng Lim 2005 T (x 1... x k 1 x k+1... x d ) = λx k, x k = 1, k [d] (1) critical points of d-linear form T j [d] x j restricted to S(m) where S(m) = S m S m d 1, S m 1 := {x R m, x = 1} C(m) := R m 1... R m d variety of rank one tensors (+zero tensor) Claim: Singular tuples of T are the critical points of dist(t, C(m)). min t R T t j [d] x j = 2 T Proj span( j [d] x j ) (T ) 2 Proj span( j [d] x j ) (T ) 2 T 2 2 = Proj span( j [d] x j ) (T ) Proj span( j [d] x j ) (T ) 2 2 Proj span( j [d] x j ) (T ) 2 = T j [d] x j for (x 1,..., x d ) S(m) dist(t, C(m)) 2 = T 2 2 max (x 1,...,x m) S(m)(T j [d] x j ) 2

4 Number of singular tuples of a generic tensor Problem: Is the number of singular values and singular vectors is finite for a generic T R m an if yes what is the number? More precisely (x 1,..., x d ) P(R m 1)... P(R m d ) is a singular tuple if T (x 1... x k 1 x k+1... x d ) = λ j x k, k [d] (2) We will consider complex sigular tuples (x 1,..., x d ) in Segre variety Σ(m, C) := P(C m 1)... P(C m d ) For real tensors and real singular tuples (2) reduces to (1) with ±λ

5 Number of complex singular tuples Number of complex singular tuples is c(m), the coefficient of t m t m d 1 in f (t, m) := d i=1 ˆt m i i t m i i ˆt i t i, where ˆt i = t t i 1 + t i t d For d = 2, c(m 1, m 2 ) = min(m 1, m 2 ) as expected: ˆt m 1 1 t m 1 1 ˆt m2 2 t m 2 2 = tm1 2 t m 1 1 t m2 1 t m 2 2 ˆt 1 t 1 ˆt 2 t 2 t 2 t 1 t 1 t 2 = ( m 1 i=1 tm 1 i 1 t i 1 Stabilization: c(m, n, p) = c(m,, n, p(m, n)) for 2 )( m 2 p p(m, n) n m 2 where p(m, n) = m + n 1 j=1 tm 2 j 2 t j 1 1 ) p(m, n) = m + n 1 boundary format case for hyperdeterminants For any m 1,..., m d 1 c(m) stabilizes for p(m 1,..., m d ) = m 1 + m m d 1 (d 2) Note this is valid also for d = 2

6 Some values of c(m, n, p) I d 1, d 2, d 3 c(d 1, d 2, d 3 ) 2, 2, 2 6 2, 2, n 8 n 3 2, 3, , 3, 4 18 n 4 2, 4, , 4, n 32 n 5 2, 5, , 5, n 50 n 6 2, m, m + 1 2m 2 3, 3, , 3, , 3, n 61 n 5 3, 4, , 4, , 4, n 148 n 6 3, 5, 5 225

7 Some values of c(m, n, p) II d 1, d 2, d 3 c(d 1, d 2, d 3 ) 3, 5, , 5, n 295 n 7 3, m, m m3 2m m 4, 4, , 4, , 4, , 4, n 480 n 7 4, 5, , 5, , 5, , 5, n 1220 n 8 5, 5, , 5, , 5, , 5, , 5, n 3881 n 9

8 Stabilization

9 An outline for computation of c(m) We construct a natural vector vector bundle E(m) on Segre variety Σ(m) := P(C m 1)... P(C m d ) At each factor P(C m i ) associate v.b. E i, dual to quotient of tautological bundle at [x i ] P(C m i ) v.b. E i [xi ] = (C m i /span(x i )) Then E(m) ([x1 ],...,[x d ]) = d i=1 E i([x i ]) Each T C m induces a section in E(m) For each ([x 1 ],..., [x d ]) Σ(m) the vector in E(m) is u(x 1,..., x d ) := d i=1 T j [d]\{i}x j E(m) ([x1 ],...,[x d ]) (x 1,..., x d ) is a singular tuple of T iff u(x 1,..., x d ) = 0 Bertini s type theorem yields: section of generic T has finite number of zeros. This number is the top Chern class of E(m)

10 Approximation of symmetric tensors: rank one at most m d := (m,..., m), R }{{} n = R m d, S(d, R n ) symmetric tensors d C k - tensors of border rank at most k Thm There exists a semi-algebraic set Q S(d, R m ), dim Q < ( ) m+d 1 d for T S(d, R n ) \ Q best rank 1-approximation unique, and symmetric Prf. 1. Banach 1939, Chen-He-Li-Zhang 2012, Friedland 2013: best rank 1-approxim. of symmetric tensor can be chosen symmetric 2. Friedland-Ottaviani: f := dist(, C 1 ) S(d, R m ). If f differentiable at T then best rank 1-approximation unique up to permutation of factors in X = x 1... x d. Use 1. to deduce X symmetric 3. Friedland-Stawiska: the set Q of symmetric tensor with not unique best rank approximation is semi-algebraic

11 Approximation of symmetric tensors: b. rank k at most N(m, d) = 1 2 (( ) m+d 3 d 2 + 2m 2) for d 3, (N(m, 3) = 3m 2 2 ) Thm For d 3, 2 k N(m, d) the semi-algebraic set of all symmetric tensors for which best border rank k approximation is unique, (denoted as P k R n ), has dimension ( ) m+d 1 d. Use Kruskal s theorem to show that a symmetric tensor of the form T = k i=1 d u i, k-as above has rank k if any min(m, k) vectors from u 1,..., u k and min(k, ( ) m+d 3 d 2 ) vectors from d 2 u 1,..., d 2 u k linearly independent Problem : Is dim(r n \ P k ) < ( ) m+d 1 d? Weaker problem: Is the best border rank k-approximation to a symmetric tensor can be chosen symmetric?

12 References 1 S. Banach, Über homogene polynome in (L 2 ), Studia Math. 7 (1938), D. Cartwright and B. Sturmfels, The number of eigenvalues of a tensor, Linear Algebra Appl. 438 (2013), B. Chen, S. He, Z. Li, and S, Zhang, Maximum block improvement and polynomial optimization, SIAM J. Optimization, 22 (2012), J. Draisma, E. Horobet, G. Ottaviani, B. Sturmfels and R.R. Thomas, The Euclidean distance degree of an algebraic variety, arxiv: S. Friedland, On tensors of border rank l in C m n l, Linear Algebra and its Applications 438 (2013),

13 References 2 S. Friedland. Best rank one approximation of real symmetric tensors can be chosen symmetric, Front. Math. China, 8 (1) (2013), S. Friedland and G. Ottaviani, The number of singular vector tuples and uniqueness of best rank one approximation of tensors, Found. Comput. Math. 2014, arxiv: S. Friedland and M. Stawiska, Best approximation on semi-algebraic sets and k-border rank approximation of symmetric tensors, arxiv: L.-H. Lim. Singular values and eigenvalues of tensors: a variational approach. Proc. IEEE International Workshop on Computational Advances in Multi-Sensor Adaptive Processing (CAMSAP 05), 1 (2005),