A Framework for Structured Linearizations of Matrix Polynomials in Various Bases

Transcription

1 A Framework for Structured Linearizations of Matrix Polynomials in Various Bases Leonardo Robol Joint work with Raf Vandebril and Paul Van Dooren, KU Leuven and Université Catholique de Louvain 09 May 2016, Structured Matrix Days in Limoges 1 / 48

2 Introduction Rational vector spaces Linearizations, dual bases and structures I will talk about different mathematical objects that play very well together: 2 / 48

3 Introduction Rational vector spaces Linearizations, dual bases and structures I will talk about different mathematical objects that play very well together: Linearizations of (matrix) polynomials. 2 / 48

4 Introduction Rational vector spaces Linearizations, dual bases and structures I will talk about different mathematical objects that play very well together: Linearizations of (matrix) polynomials. Dual (minimal) bases, related to linear system theory. 2 / 48

5 Introduction Rational vector spaces Linearizations, dual bases and structures I will talk about different mathematical objects that play very well together: Linearizations of (matrix) polynomials. Dual (minimal) bases, related to linear system theory. How to take structures from polynomials and preserve them in linearizations (spoiler alert: using dual bases!). 2 / 48

6 Introduction Rational vector spaces Linearizations, dual bases and structures I will talk about different mathematical objects that play very well together: Linearizations of (matrix) polynomials. Dual (minimal) bases, related to linear system theory. How to take structures from polynomials and preserve them in linearizations (spoiler alert: using dual bases!). How to deal with different bases in the definition of a polynomial. 2 / 48

7 Introduction Rational vector spaces Vector spaces of rational functions One can consider the vector space F n (λ) of rational functions in the variable λ: v(λ) F n (λ) v(λ) = p 1 (λ) q 1 (λ). p n(λ) q n(λ) 3 / 48

8 Introduction Rational vector spaces Vector spaces of rational functions One can consider the vector space F n (λ) of rational functions in the variable λ: v(λ) F n (λ) v(λ) = p 1 (λ) q 1 (λ). p n(λ) q n(λ) F(λ) is a field, so everything is nice. However, every vector can be written as v(λ) = q(λ) 1 ṽ(λ) with v(λ) F[λ]. 3 / 48

9 Introduction Rational vector spaces Vector spaces of rational functions One can consider the vector space F n (λ) of rational functions in the variable λ: v(λ) F n (λ) v(λ) = p 1 (λ) q 1 (λ). p n(λ) q n(λ) F(λ) is a field, so everything is nice. However, every vector can be written as v(λ) = q(λ) 1 ṽ(λ) with v(λ) F[λ]. Switching between polynomials and rational functions is useful for theoretical reasons. 3 / 48

10 Introduction Rational vector spaces Bases for rational vector spaces Let V F n (λ) a r-dimensional subspace of the vector space on the field of rational functions. 4 / 48

11 Introduction Rational vector spaces Bases for rational vector spaces Let V F n (λ) a r-dimensional subspace of the vector space on the field of rational functions. We can write a basis for V as A(λ) F n r (λ), so that such that: f (λ) V f (λ) = A(λ)w f (λ) 4 / 48

12 Introduction Rational vector spaces Bases for rational vector spaces Let V F n (λ) a r-dimensional subspace of the vector space on the field of rational functions. We can write a basis for V as A(λ) F n r (λ), so that such that: f (λ) V f (λ) = A(λ)w f (λ) V admits a basis composed only of polynomials. 4 / 48

13 Introduction Rational vector spaces Bases for rational vector spaces Let V F n (λ) a r-dimensional subspace of the vector space on the field of rational functions. We can write a basis for V as A(λ) F n r (λ), so that such that: f (λ) V f (λ) = A(λ)w f (λ) V admits a basis composed only of polynomials. Among these bases, we can look for the ones whose sum of column-degrees is minimal. The basis is not unique, but its column-degrees are. We call these bases minimal, and their column degrees minimal indices. 4 / 48

14 Introduction Dual bases Dual spaces We can look for the dual space V, which has dimension n r. V = {g(λ) F n (λ) g(λ) T f (λ) = 0 f (λ) V}. 5 / 48

15 Introduction Dual bases Dual spaces We can look for the dual space V, which has dimension n r. V = {g(λ) F n (λ) g(λ) T f (λ) = 0 f (λ) V}. If B(λ) is a dual basis to A(λ) if A(λ) T B(λ) = 0. This is equivalent to saying that B(λ) is a basis of V. If both A(λ) and B(λ) are minimal we say that they form a pair of dual minimal bases. 5 / 48

16 Introduction Dual bases Dual spaces We can look for the dual space V, which has dimension n r. V = {g(λ) F n (λ) g(λ) T f (λ) = 0 f (λ) V}. If B(λ) is a dual basis to A(λ) if A(λ) T B(λ) = 0. This is equivalent to saying that B(λ) is a basis of V. If both A(λ) and B(λ) are minimal we say that they form a pair of dual minimal bases. 1 λ λ λ k. 1 = 0 are dual minimal bases. 5 / 48

17 Introduction Dual bases Properties of dual minimal bases Let A(λ) T B(λ) = 0 be a pair of dual minimal bases. The sum of degrees of a minimal basis and of a minimal basis of the dual space always coincide: r n r deg A(λ)e i = B(λ)e i. i=1 i=1 6 / 48

18 Introduction Dual bases Properties of dual minimal bases Let A(λ) T B(λ) = 0 be a pair of dual minimal bases. The sum of degrees of a minimal basis and of a minimal basis of the dual space always coincide: r n r deg A(λ)e i = B(λ)e i. i=1 i=1 Pairs of minimal bases behave in a good way under perturbations: 6 / 48

19 Introduction Dual bases Properties of dual minimal bases Let A(λ) T B(λ) = 0 be a pair of dual minimal bases. The sum of degrees of a minimal basis and of a minimal basis of the dual space always coincide: r n r deg A(λ)e i = B(λ)e i. i=1 i=1 Pairs of minimal bases behave in a good way under perturbations: Minimality is preserved for small enough changes. 6 / 48

20 Introduction Dual bases Properties of dual minimal bases Let A(λ) T B(λ) = 0 be a pair of dual minimal bases. The sum of degrees of a minimal basis and of a minimal basis of the dual space always coincide: r n r deg A(λ)e i = B(λ)e i. i=1 Pairs of minimal bases behave in a good way under perturbations: Minimality is preserved for small enough changes. Small perturbations to A(λ) correspond to small perturbations of the dual space (in an appropriate sense). i=1 6 / 48

21 Introduction Dual bases Properties of dual minimal bases Let A(λ) T B(λ) = 0 be a pair of dual minimal bases. The sum of degrees of a minimal basis and of a minimal basis of the dual space always coincide: r n r deg A(λ)e i = B(λ)e i. i=1 Pairs of minimal bases behave in a good way under perturbations: Minimality is preserved for small enough changes. Small perturbations to A(λ) correspond to small perturbations of the dual space (in an appropriate sense). i=1 These facts are crucial in proving backward stability results. 6 / 48

22 Linearizations Basic examples Building linearizations We are interested in computing eigenvalues and eigenstructure of a matrix polynomial P(λ): P(λ) = n P i λ i, or, more generally, P(λ) = i=0 n P i φ i (λ). i=0 where Φ = {φ 0 (λ),..., φ n (λ)} is some polynomial basis (monomials, Chebyshev, Lagrange, Newton,... ). 7 / 48

23 Linearizations Basic examples Building linearizations We are interested in computing eigenvalues and eigenstructure of a matrix polynomial P(λ): P(λ) = n P i λ i, or, more generally, P(λ) = i=0 n P i φ i (λ). i=0 where Φ = {φ 0 (λ),..., φ n (λ)} is some polynomial basis (monomials, Chebyshev, Lagrange, Newton,... ). We look for a pencil L(λ) with the same spectral properties. We say that L(λ) is a linearization for P(λ) if and only if: P(λ) I m(n 1) = E(λ)L(λ)F (λ), det E(λ), det F (λ) invertible in F[λ]. 7 / 48

24 Linearizations Basic examples From dual bases to linearizations Given a basis Φ := {φ 0 (λ),..., φ n (λ)} we say that L φ (λ) and π φ (λ) are dual bases associated to Φ if: φ n (λ) L φ (λ) T π φ (λ) = 0, π φ (λ) =. φ 0 (λ) Theorem If L φ (λ) and π φ (λ) are dual bases associated to Φ then [ W T ] W n (λ) T (λ) L(λ) := L T φ (λ) I, W (λ) =. m W 0 (λ) T linearizes P(λ) := W T (λ)(π φ (λ) I m ) = n i=0 W i(λ)φ i (λ). 8 / 48

25 Linearizations Basic examples Almost a proof We check what the eigenvalues of such a pencil are: { W T (λ)v = 0 L(λ)v = 0 (L T (λ) I m )v = 0 9 / 48

26 Linearizations Basic examples Almost a proof We check what the eigenvalues of such a pencil are: { W T (λ)v = 0 L(λ)v = 0 (L T (λ) I m )v = 0 Since v is in the dual of L(λ) we can write it as v = (π(λ) I m )w v where w v 0 so that: W T (λ)v = 0 W T (λ)(π(λ) I m )w v = 0 P(λ)w v = 0. 9 / 48

27 Linearizations Basic examples Almost a proof We check what the eigenvalues of such a pencil are: { W T (λ)v = 0 L(λ)v = 0 (L T (λ) I m )v = 0 Since v is in the dual of L(λ) we can write it as v = (π(λ) I m )w v where w v 0 so that: W T (λ)v = 0 W T (λ)(π(λ) I m )w v = 0 P(λ)w v = 0. Basic idea: The eigenvector must lie in the dual of L(λ), thus has the correct structure. 9 / 48

28 since p(λ) = [ λp n + p n 1... p 0 ] π(λ). 10 / 48 Structured Linearizations Linearizations Basic examples A well-known example One can consider the pair of dual minimal bases 1 L(λ) = λ , π(λ) = λ λ n 1 so that λp n + p n p 0 1 λ n L(λ) = linearizes p(λ) := p i λ i. i=0 1 λ. λ 1

29 Linearizations Basic examples Other well-known examples... the previous approach allows for linearization in arbitrary bases! λp n + p n 1 p n 2 p n... p 1 p 0 1 2λ 1 L(λ) = λ 1 1 λ is a linearization for the polynomial expressed in the Chebyshev basis of the first kind: n p(λ) = p i T i (λ). i=0 11 / 48

30 Linearizations Basic examples Other well-known examples... the previous approach allows for linearization in arbitrary bases! λp n + p n 1 p n 2 p n... p 1 p 0 1 2λ 1 L(λ) = λ 1 1 λ is a linearization for the polynomial expressed in the Chebyshev basis of the first kind: n p(λ) = p i T i (λ). i=0 Already in [Amirlaslani, Corless, Lancaster, 2009]: all the orthogonal bases and also Newton, Lagrange, Hermite are possible. 11 / 48

31 Linearizations Basic examples Just to get the idea Assume we have an orthogonal basis φ i (λ) satisfying the relation: αφ j+1 (λ) = (λ β)φ j (λ) γφ j 1 (λ), α 0, j > 0, (1) 12 / 48

32 Linearizations Basic examples Just to get the idea Assume we have an orthogonal basis φ i (λ) satisfying the relation: αφ j+1 (λ) = (λ β)φ j (λ) γφ j 1 (λ), α 0, j > 0, (1) Then α (β λ) γ L φ (λ) T := α (β λ) γ φ 0 (λ) φ 1 (λ) is a basis dual to π φ (λ). 12 / 48

33 Linearizations Basic examples Using two dual bases at the same time The construction of linearizations can be generalized to the case where we have two pairs of dual minimal bases. 13 / 48

34 Linearizations Basic examples Using two dual bases at the same time The construction of linearizations can be generalized to the case where we have two pairs of dual minimal bases. Let Φ, Ψ be two polynomial bases, and L φ (λ), π φ (λ) and L ψ (λ), π ψ (λ) the corresponding dual minimal bases. 13 / 48

35 Linearizations Basic examples Using two dual bases at the same time The construction of linearizations can be generalized to the case where we have two pairs of dual minimal bases. Let Φ, Ψ be two polynomial bases, and L φ (λ), π φ (λ) and L ψ (λ), π ψ (λ) the corresponding dual minimal bases. Theorem The matrix pencil L(λ) := [ λm1 + M 0 ] L φ (λ) L ψ (λ) T 0 is a linearization for the polynomial p(λ) := π ψ (λ) T (λm 1 + M 0 )π φ (λ). 13 / 48

36 Linearizations Fiedler and more bases Relation with Fiedler linearizations The family of linearizations that can be built using the previous result is very large: 14 / 48

37 Linearizations Fiedler and more bases Relation with Fiedler linearizations The family of linearizations that can be built using the previous result is very large: The Frobenius linearizations (the classical companions) are a particular case. 14 / 48

38 Linearizations Fiedler and more bases Relation with Fiedler linearizations The family of linearizations that can be built using the previous result is very large: The Frobenius linearizations (the classical companions) are a particular case. Every Fiedler pencil/companion is just a permutation of a pencil in the previous form. 14 / 48

39 Linearizations Fiedler and more bases Relation with Fiedler linearizations The family of linearizations that can be built using the previous result is very large: The Frobenius linearizations (the classical companions) are a particular case. Every Fiedler pencil/companion is just a permutation of a pencil in the previous form. However, many more linearizations fits in this class! 14 / 48

40 Linearizations Fiedler and more bases Equivalence to Fiedler linearizations Consider p(λ) = p 4 λ 4 + p 3 λ 3 + p 2 λ 2 + p 1 λ + p 0. A Fiedler linearization looks like: λp 4 + p λb A = p 2 λ p 1 p λ λ 15 / 48

41 Linearizations Fiedler and more bases Equivalence to Fiedler linearizations Consider p(λ) = p 4 λ 4 + p 3 λ 3 + p 2 λ 2 + p 1 λ + p 0. A Fiedler linearization looks like: λp 4 + p λb A = p 2 λ p 1 p λ λ Permuting the above pencil in a good way yields: λp 4 + p p 2 p 1 p 0 λ 1 λ λ 0 [ ] = λm1 + M 0 L φ (λ) L ψ (λ) T 0 15 / 48

42 Linearizations Fiedler and more bases Exploiting the extra freedom According to the previous Theorem, we can write the linearized polynomial: λp 4 + p p 2 p 1 p 0 λ 1 λ λ 0 [ λ 1 ] [ ] λp 4 + p 3 0 λ 2 λ. p 2 p 1 p / 48

43 Linearizations Fiedler and more bases Exploiting the extra freedom According to the previous Theorem, we can write the linearized polynomial: λp 4 + p p 2 p 1 p 0 λ 1 λ λ 0 [ λ 1 ] [ ] λp 4 + p 3 0 λ 2 λ. p 2 p 1 p 0 1 The latter product corresponds to the sum of all the elements in the Hadamard product: [ ] [ λp4 + p 3 λ 3 λ 2 ] λ p 2 p 1 p 0 λ 2 λ 1 which clearly is p(λ). However, reshuffling is allowed! 16 / 48

44 Linearizations Fiedler and more bases Exploiting the extra freedom We can get symmetric linearizations for odd degree polynomials: λp 5 + p 4 1 λp 3 + p 2 λ 1 λp 1 + p 0 λ 1 λ λ 0 0 Which trivially extends to block symmetric linearizations for matrix polynomials, which are symmetric whenever the polynomial is. 17 / 48

45 Linearizations Fiedler and more bases Exploiting the extra freedom We can get symmetric linearizations for odd degree polynomials: λp 5 + p 4 1 λp 3 + p 2 λ 1 λp 1 + p 0 λ 1 λ λ 0 0 Which trivially extends to block symmetric linearizations for matrix polynomials, which are symmetric whenever the polynomial is. Main underlying idea: Construct the two bases L φ (λ) and L ψ (λ) so that they have the same symmetry of the matrix polynomial, and you will get a structured linearization. 17 / 48

46 Preserving structures Symmetries Why do we care about symmetries? 18 / 48

47 Preserving structures Symmetries Why do we care about symmetries? They are beautiful. 18 / 48

48 Preserving structures Symmetries Why do we care about symmetries? They are beautiful. They are fun to work with. 18 / 48

49 Preserving structures Symmetries Why do we care about symmetries? They are beautiful. They are fun to work with. But since we need to get research funded we also notice that: 18 / 48

50 Preserving structures Symmetries Why do we care about symmetries? They are beautiful. They are fun to work with. But since we need to get research funded we also notice that: They are naturally present in many engineering applications. 18 / 48

51 Preserving structures Symmetries Why do we care about symmetries? They are beautiful. They are fun to work with. But since we need to get research funded we also notice that: They are naturally present in many engineering applications. Preserving symmetries in the linearizations means preserving also spectral symmetries. 18 / 48

52 Preserving structures Symmetries Why do we care about symmetries? They are beautiful. They are fun to work with. But since we need to get research funded we also notice that: They are naturally present in many engineering applications. Preserving symmetries in the linearizations means preserving also spectral symmetries. Structured methods can give us approximations to the eigenvalues that make sense in the underlying model and have a lower condition number (structured vs unstructured). 18 / 48

53 Preserving structures Symmetries Why do we care about symmetries? They are beautiful. They are fun to work with. But since we need to get research funded we also notice that: They are naturally present in many engineering applications. Preserving symmetries in the linearizations means preserving also spectral symmetries. Structured methods can give us approximations to the eigenvalues that make sense in the underlying model and have a lower condition number (structured vs unstructured). We get better results in the end. 18 / 48

54 Preserving structures Symmetries Why do we care about symmetries? They are beautiful. They are fun to work with. But since we need to get research funded we also notice that: They are naturally present in many engineering applications. Preserving symmetries in the linearizations means preserving also spectral symmetries. Structured methods can give us approximations to the eigenvalues that make sense in the underlying model and have a lower condition number (structured vs unstructured). We get better results in the end. Famous example: simulation of vibrations for high speed trains: palindromic matrix polynomials. 18 / 48

55 Preserving structures Symmetries Matrix and scalar polynomials Most of the things that you see here for scalar polynomials are indeed true (and useful) for matrix polynomials. 19 / 48

56 Preserving structures Symmetries Matrix and scalar polynomials Most of the things that you see here for scalar polynomials are indeed true (and useful) for matrix polynomials. Just take some I m and attach them on the right of (almost) every matrix you can see, that is [ ] λm1 + M 0 L φ (λ) L T ψ (λ) 0 are linearizations for: [ λm1 + M 0 ] L φ (λ) I m L T ψ (λ) I m 0 π φ (λ) T (λm 1 +M 0 )π ψ (λ), (π φ (λ) T I m )(λm 1 +M 0 )(π ψ (λ) I m ) 19 / 48

57 Preserving structures Symmetries Matrix and scalar polynomials Most of the things that you see here for scalar polynomials are indeed true (and useful) for matrix polynomials. Just take some I m and attach them on the right of (almost) every matrix you can see, that is [ ] λm1 + M 0 L φ (λ) L T ψ (λ) 0 are linearizations for: [ λm1 + M 0 ] L φ (λ) I m L T ψ (λ) I m 0 π φ (λ) T (λm 1 +M 0 )π ψ (λ), (π φ (λ) T I m )(λm 1 +M 0 )(π ψ (λ) I m ) I will switch back and forth between the two notations. 19 / 48

58 Preserving structures Even odd case Even/odd matrix polynomials Another common structure found in applications is -even/odd polynomial ( {T,, 1}). 20 / 48

59 Preserving structures Even odd case Even/odd matrix polynomials Another common structure found in applications is -even/odd polynomial ( {T,, 1}). P(λ) = P( λ) P(λ) = P( λ) even odd 20 / 48

60 Preserving structures Even odd case Even/odd matrix polynomials Another common structure found in applications is -even/odd polynomial ( {T,, 1}). P(λ) = P( λ) P(λ) = P( λ) even odd Symmetries are reflect in the spectrum: { λ eigenvalue λ eigenvalue {T, 1} λ eigenvalue λ eigenvalue = 20 / 48

61 Preserving structures Even odd case Even/odd matrix polynomials Another common structure found in applications is -even/odd polynomial ( {T,, 1}). P(λ) = P( λ) P(λ) = P( λ) even odd Symmetries are reflect in the spectrum: { λ eigenvalue λ eigenvalue {T, 1} λ eigenvalue λ eigenvalue = Given such a polynomial, how to construct a -even/odd linearization? Let s construct two adapted L(λ) / 48

62 Preserving structures Even odd case -even/odd linearizations! 1 λ 1 λ L φ (λ) T =......, L ψ (λ) T = λ 1 λ Notice that L φ (λ) = L ψ ( λ), so... L(λ) = [ ] λm1 + M 0 L φ (λ) I m L T ψ (λ) I, M m 0 1 = M1 T, M 0 = M0 T is a T -even linearization for the matrix polynomial: P(λ) = (π T φ (λ) I m)(λm 1 + M 0 )(π ψ (λ) I m ). 21 / 48

63 Preserving structures Even odd case The final step We can easily check that: λ n 1 ( 1) n 1 λ n 1 π φ (λ) =. λ, π ψ(λ) =. λ, / 48

64 Preserving structures Even odd case The final step We can easily check that: λ n 1 ( 1) n 1 λ n 1 π φ (λ) =. λ, π ψ(λ) =. λ, 1 1 and that a reasonable choice for λm 1 + M 0 is given by: ( 1) n 1 (λp 2n 1 + P 2n 2 ) λm 1 + M 0 =.... P0 + λp1 If P(λ) is -even the λm 1 + M 0 is also -even and so we have built a structured linearization. 22 / 48

65 Preserving structures Palindromic linearizations Palindromic matrix polynomials Another interesting case: P(λ) = rev P(λ), P(λ) = rev P(λ). Here rev P(λ) := λ deg P P(λ 1 ). This induces the spectral symmetry { λ eigenvalue λ 1 eigenvalue {T, 1} λ eigenvalue λ 1 eigenvalue = The = 1 case is not interesting, since it is impossible to build structured linearizations. In the other cases, however, we can use more or less the same trick of before. 23 / 48

66 Preserving structures Palindromic linearizations Building palindromic dual bases 1 λ λ L φ (λ) T =......, L ψ (λ) T = 1 λ λ 1 24 / 48

67 Preserving structures Palindromic linearizations Building palindromic dual bases 1 λ λ L φ (λ) T =......, L ψ (λ) T = 1 λ We follow the now usual approach and we have: λ n 1 1 λ π φ (λ) =. λ 1, π ψ(λ) =. λ n 1, λ 1 24 / 48

68 Preserving structures Palindromic linearizations... and then a palindromic linearization When building the linearization we have the correct symmetries, since rev L φ (λ) = L ψ (λ), and so L(λ) = is a palindromic linearization. [ λm + M ] L φ (λ) L ψ (λ) T 0 25 / 48

69 Preserving structures Palindromic linearizations... and then a palindromic linearization When building the linearization we have the correct symmetries, since rev L φ (λ) = L ψ (λ), and so L(λ) = [ λm + M ] L φ (λ) L ψ (λ) T 0 is a palindromic linearization. How to choose M? Some computations show that 0 m 0 m P 0 M =... 0 m 0 m Pn 1 is the right choice for a palindromic polynomial of degree 2n / 48

70 Preserving structures Palindromic linearizations Midway conclusions We are considering a family of matrices that includes Fiedler linearizations, and we have found structured linearizations which are also possible in the Fielder framework. 26 / 48

71 Preserving structures Palindromic linearizations Midway conclusions We are considering a family of matrices that includes Fiedler linearizations, and we have found structured linearizations which are also possible in the Fielder framework. However, we build bases which reflect the structure of the problem, instead of fighting with permutations and combinatorics. 26 / 48

72 Preserving structures Palindromic linearizations Midway conclusions We are considering a family of matrices that includes Fiedler linearizations, and we have found structured linearizations which are also possible in the Fielder framework. However, we build bases which reflect the structure of the problem, instead of fighting with permutations and combinatorics. I think it is much easier to go this way. 26 / 48

73 Preserving structures Palindromic linearizations Midway conclusions We are considering a family of matrices that includes Fiedler linearizations, and we have found structured linearizations which are also possible in the Fielder framework. However, we build bases which reflect the structure of the problem, instead of fighting with permutations and combinatorics. I think it is much easier to go this way. Can we do other interesting things with this framework? 26 / 48

74 Preserving structures Palindromic linearizations Midway conclusions We are considering a family of matrices that includes Fiedler linearizations, and we have found structured linearizations which are also possible in the Fielder framework. However, we build bases which reflect the structure of the problem, instead of fighting with permutations and combinatorics. I think it is much easier to go this way. Can we do other interesting things with this framework? Yes, let s see something completely different! 26 / 48

75 Intersections of polynomials Scalar polynomials Intersection of polynomials Consider the problem: find the values of λ such that with p 1 (λ), p 2 (λ) polynomials. p 1 (λ) = p 2 (λ) 27 / 48

76 Intersections of polynomials Scalar polynomials Intersection of polynomials Consider the problem: find the values of λ such that p 1 (λ) = p 2 (λ) with p 1 (λ), p 2 (λ) polynomials. That s essentially saying: find the roots of p 1 (λ) p 2 (λ). 27 / 48

77 Intersections of polynomials Scalar polynomials Intersection of polynomials Consider the problem: find the values of λ such that p 1 (λ) = p 2 (λ) with p 1 (λ), p 2 (λ) polynomials. That s essentially saying: find the roots of p 1 (λ) p 2 (λ). However, what to do if p 1 (λ) and p 2 (λ) are represented using different bases? Can we still easily solve the problem? 27 / 48

78 Intersections of polynomials Scalar polynomials Intersection of polynomials Consider the problem: find the values of λ such that p 1 (λ) = p 2 (λ) with p 1 (λ), p 2 (λ) polynomials. That s essentially saying: find the roots of p 1 (λ) p 2 (λ). However, what to do if p 1 (λ) and p 2 (λ) are represented using different bases? Can we still easily solve the problem? Convert them to the same basis, and use a companion matrix for that basis. Use our framework in a creative way. 27 / 48

79 Intersections of polynomials Scalar polynomials Let s recap L(λ) = [ λm1 + M 0 ] L φ (λ) L ψ (λ) T 0 is a linearization for p(λ) := π T φ (λ)(λm 1 + M 0 )π ψ (λ). 28 / 48

80 Intersections of polynomials Scalar polynomials Let s recap L(λ) = [ λm1 + M 0 ] L φ (λ) L ψ (λ) T 0 is a linearization for p(λ) := π T φ (λ)(λm 1 + M 0 )π ψ (λ). Let w φ, w ψ constant vectors such that w T φ π φ(λ) = 1 = w T ψ π ψ(λ) (the coordinates of the constant 1 in our basis). 28 / 48

81 Intersections of polynomials Scalar polynomials Let s recap L(λ) = [ λm1 + M 0 ] L φ (λ) L ψ (λ) T 0 is a linearization for p(λ) := π T φ (λ)(λm 1 + M 0 )π ψ (λ). Let w φ, w ψ constant vectors such that w T φ π φ(λ) = 1 = w T ψ π ψ(λ) (the coordinates of the constant 1 in our basis). If we set λm 1 + M 0 = w φ p T 1 p 2w T ψ we have: p(λ) = p T 1 π ψ (λ) π T φ (λ)p 2 = η ɛ p 1,i ψ i (λ) p 2,i φ i (λ), i=0 i=0 assuming η and ɛ are the number of columns of L ψ (λ) and L φ (λ). 28 / 48

82 Intersections of polynomials Scalar polynomials What is the advantage? We have linearized a difference of polynomials expressed in different bases directly (no computations needed!). 29 / 48

83 Intersections of polynomials Scalar polynomials What is the advantage? We have linearized a difference of polynomials expressed in different bases directly (no computations needed!). The linearizations directly contain the coefficients of the two polynomials, thus avoiding the potentially ill-conditioned basis change. 29 / 48

84 Intersections of polynomials Scalar polynomials What is the advantage? We have linearized a difference of polynomials expressed in different bases directly (no computations needed!). The linearizations directly contain the coefficients of the two polynomials, thus avoiding the potentially ill-conditioned basis change. However, L(λ) is a linearization for a polynomial of grade ɛ + η, while we have degree max{ɛ, η}. This gives us many infinite eigenvalues. 29 / 48

85 Intersections of polynomials Scalar polynomials Infinite eigenvalues Infinite eigenvalues can be deflated a posteriori. 30 / 48

86 Intersections of polynomials Scalar polynomials Infinite eigenvalues Infinite eigenvalues can be deflated a posteriori. We can easily characterize their complete eigenstructure. They form a very long Jordan chain at infinity very badly conditioned! 30 / 48

87 Intersections of polynomials Scalar polynomials Infinite eigenvalues Infinite eigenvalues can be deflated a posteriori. We can easily characterize their complete eigenstructure. They form a very long Jordan chain at infinity very badly conditioned! However, the others eigenvalues can be perfectly conditioned. 30 / 48

88 Intersections of polynomials Scalar polynomials Infinite eigenvalues Infinite eigenvalues can be deflated a posteriori. We can easily characterize their complete eigenstructure. They form a very long Jordan chain at infinity very badly conditioned! However, the others eigenvalues can be perfectly conditioned. We can also deflate the infinite eigenstructure by the staircase algorithm (since we know the eigenstructure, no rank decision are needed, and the approach is perfectly stable). 30 / 48

89 Intersections of polynomials Scalar polynomials Numerical accuracy I claimed that this approach is more stable than conversion to a common basis. Is this true? 31 / 48

90 Intersections of polynomials Scalar polynomials Numerical accuracy I claimed that this approach is more stable than conversion to a common basis. Is this true? Norm of absolute errors Monomial Chebyshev Linearization Lin + Deflation Degree 31 / 48

91 Intersections of polynomials Going to rational functions Extensions to rational functions The same idea can be applied to a slightly more general problem. Find the solutions of: f (λ) := p(λ) q(λ) + r(λ) s(λ) = 0, with p(λ), q(λ), r(λ), and s(λ) polynomials. 32 / 48

92 Intersections of polynomials Going to rational functions Extensions to rational functions The same idea can be applied to a slightly more general problem. Find the solutions of: f (λ) := p(λ) q(λ) + r(λ) s(λ) = 0, with p(λ), q(λ), r(λ), and s(λ) polynomials. This is equivalent to finding the roots of t(λ) := p(λ)s(λ) + q(λ)r(λ). 32 / 48

93 Intersections of polynomials Going to rational functions A linearization for rational functions Similar to the previous case, we have: [ ps L(λ) = T + qr T ] L φ (λ) L ψ (λ) T 0 which is a linearization for t(λ). The eigenvalues are solutions of f (λ) := p(λ) q(λ) + r(λ) s(λ) = 0, where p(λ), q(λ) are expressed in the Φ basis, and r(λ) and s(λ) in the Ψ basis. 33 / 48

94 Intersections of polynomials Going to rational functions Easy proof We can easily check what is linearized by L(λ): t(λ) = π T φ (λ)(pst + qr T )π ψ (λ) = = p(λ)s(λ) + q(λ)r(λ). where p(λ), q(λ) are represented in the basis Φ, and r(λ), s(λ) in the basis Ψ. 34 / 48

95 Intersections of polynomials Going to rational functions Easy proof We can easily check what is linearized by L(λ): t(λ) = π T φ (λ)(pst + qr T )π ψ (λ) = = p(λ)s(λ) + q(λ)r(λ). where p(λ), q(λ) are represented in the basis Φ, and r(λ), s(λ) in the basis Ψ. Notice the the top-left rank 2 block has no term depending on λ. 34 / 48

96 Intersections of polynomials Going to rational functions Good news Under mild hypothesis on degree matching we have only 1 simple infinite eigenvalue. 35 / 48

97 Intersections of polynomials Going to rational functions Good news Under mild hypothesis on degree matching we have only 1 simple infinite eigenvalue. If one of the two basis satisfies the relation π ɛ,φ (λ) = (λa + B)π ɛ 1,φ (λ) we can get rid of all the infinite eigenvalues by slightly adjusting the construction. 35 / 48

98 Intersections of polynomials Going to rational functions Good news Under mild hypothesis on degree matching we have only 1 simple infinite eigenvalue. If one of the two basis satisfies the relation π ɛ,φ (λ) = (λa + B)π ɛ 1,φ (λ) we can get rid of all the infinite eigenvalues by slightly adjusting the construction. Hint: any basis you can think of satisfies the above property (monomial, all orthogonal, Newton, Lagrange, Hermite are all ok). 35 / 48

99 Intersections of polynomials Going to rational functions Another way to write p(λ)... L(λ) = [ M(λ) ] Lφ (λ) L ψ (λ) T 0 linearizes M i,j (λ)ψ i (λ)φ j (λ). i,j 36 / 48

100 Intersections of polynomials Going to rational functions Another way to write p(λ)... L(λ) = [ M(λ) ] Lφ (λ) L ψ (λ) T 0 linearizes M i,j (λ)ψ i (λ)φ j (λ). i,j We can see this as a linearization in the product family ψ φ := {φ i (λ)ψ j (λ) i = 1,..., ɛ, j = 1,..., η} 36 / 48

101 Intersections of polynomials Going to rational functions Another way to write p(λ)... L(λ) = [ M(λ) ] Lφ (λ) L ψ (λ) T 0 linearizes M i,j (λ)ψ i (λ)φ j (λ). i,j We can see this as a linearization in the product family ψ φ := {φ i (λ)ψ j (λ) i = 1,..., ɛ, j = 1,..., η} Not a basis in general. 36 / 48

102 Intersections of polynomials Going to rational functions Another way to write p(λ)... L(λ) = [ M(λ) ] Lφ (λ) L ψ (λ) T 0 linearizes M i,j (λ)ψ i (λ)φ j (λ). i,j We can see this as a linearization in the product family ψ φ := {φ i (λ)ψ j (λ) i = 1,..., ɛ, j = 1,..., η} Not a basis in general. However, when φ i and ψ j this is a generating family for the polynomials of degree up to ɛ + η. 36 / 48

103 Intersections of polynomials Going to rational functions Another way to write p(λ)... L(λ) = [ M(λ) ] Lφ (λ) L ψ (λ) T 0 linearizes M i,j (λ)ψ i (λ)φ j (λ). i,j We can see this as a linearization in the product family ψ φ := {φ i (λ)ψ j (λ) i = 1,..., ɛ, j = 1,..., η} Not a basis in general. However, when φ i and ψ j this is a generating family for the polynomials of degree up to ɛ + η. We can exploit the redundancy in a good way. 36 / 48

104 Intersections of polynomials More and more bases!... which leads to a natural question Can we handle more than two bases? 37 / 48

105 Intersections of polynomials More and more bases!... which leads to a natural question Can we handle more than two bases? Given φ (1),..., φ (j) polynomial bases I can construct their product: φ (1)... φ (j) 37 / 48

106 Intersections of polynomials More and more bases!... which leads to a natural question Can we handle more than two bases? Given φ (1),..., φ (j) polynomial bases I can construct their product: φ (1)... φ (j) We know how to handle j = 1 (the classical companion case). 37 / 48

107 Intersections of polynomials More and more bases!... which leads to a natural question Can we handle more than two bases? Given φ (1),..., φ (j) polynomial bases I can construct their product: φ (1)... φ (j) We know how to handle j = 1 (the classical companion case).... and also j = 2 (what we have seen until now). 37 / 48

108 Intersections of polynomials More and more bases!... which leads to a natural question Can we handle more than two bases? Given φ (1),..., φ (j) polynomial bases I can construct their product: φ (1)... φ (j) We know how to handle j = 1 (the classical companion case).... and also j = 2 (what we have seen until now). What about going higher? 37 / 48

109 Intersections of polynomials More and more bases! Enlarged dual bases We can deal with this by constructing enlarged dual bases. φ 0 (λ)ψ 0 (λ). π φ ψ (λ) := φ i (λ)ψ j (λ). φ ɛ (λ)ψ η (λ) 38 / 48

110 Intersections of polynomials More and more bases! Enlarged dual bases We can deal with this by constructing enlarged dual bases. φ 0 (λ)ψ 0 (λ). π φ ψ (λ) := φ i (λ)ψ j (λ). φ ɛ (λ)ψ η (λ) Can we build a suitable L(λ) such that L(λ) T π φ ψ (λ) = 0? 38 / 48

111 Intersections of polynomials More and more bases! Enlarged dual bases We can deal with this by constructing enlarged dual bases. φ 0 (λ)ψ 0 (λ). π φ ψ (λ) := φ i (λ)ψ j (λ). φ ɛ (λ)ψ η (λ) Can we build a suitable L(λ) such that L(λ) T π φ ψ (λ) = 0? Yes! 38 / 48

112 Intersections of polynomials More and more bases! Product dual bases [ L k,φ ψ (λ) T A Lη,ψ (λ) = T ] L ɛ,φ (λ) T w T, k := (ɛ + 1)(η + 1) 1, form a dual basis together with π φ ψ (λ), where A is any invertible matrix and w any vector such that w T π η,ψ (λ) is a nonzero constant. 39 / 48

113 Intersections of polynomials More and more bases! Product dual bases [ L k,φ ψ (λ) T A Lη,ψ (λ) = T ] L ɛ,φ (λ) T w T, k := (ɛ + 1)(η + 1) 1, form a dual basis together with π φ ψ (λ), where A is any invertible matrix and w any vector such that w T π η,ψ (λ) is a nonzero constant. We can just plug this ingredient in our framework and obtain the desired linearizations! 39 / 48

114 Intersections of polynomials More and more bases! Some examples Let p(λ) = 3 i=0 p iλ i a degree 3 polynomial. 40 / 48

115 Intersections of polynomials More and more bases! Some examples Let p(λ) = 3 i=0 p iλ i a degree 3 polynomial. Choosing {ψ i } = {1, λ, λ 2 } and {φ i } = {1} yields Frobenius form: λp 3 + p 2 p 1 p 0 L(λ) = 1 λ. 1 λ 40 / 48

116 Intersections of polynomials More and more bases! Some examples Let p(λ) = 3 i=0 p iλ i a degree 3 polynomial. Choosing {ψ i } = {1, λ, λ 2 } and {φ i } = {1} yields Frobenius form: λp 3 + p 2 p 1 p 0 L(λ) = 1 λ. 1 λ Choosing {ψ i } = {φ i } = {1, λ} yields, e.g., the symmetric 1 λp 3 + p 2 2 p 1 1 L(λ) = 1 2 p 1 p 0 λ. 1 λ 0 40 / 48

117 Intersections of polynomials More and more bases! Some examples Choosing {ψ i } = {1, λ} {1, λ} and {φ i } = {1} yields L(λ) = 1 λp 3 + p 2 2 p 1 1 λ 1 2 p 1 p 0 1 λ 1 λ. 41 / 48

118 Intersections of polynomials More and more bases! Some examples Choosing {ψ i } = {1, λ} {1, λ} and {φ i } = {1} yields L(λ) = 1 λp 3 + p 2 2 p 1 1 λ 1 2 p 1 p 0 1 λ 1 λ. Note, the dimension could increase (not strong anymore). 41 / 48

119 Future outlook Backward stability Backward stability Work ongoing in collaboration with many people (Piers Lawrence, Françoise Tisseur, Raf Vandebril, Paul Van Dooren,... ). 42 / 48

120 Future outlook Backward stability Backward stability Work ongoing in collaboration with many people (Piers Lawrence, Françoise Tisseur, Raf Vandebril, Paul Van Dooren,... ). I want to solve the eigenvalue problem L(λ) via QZ, with [ ] W T L(λ) = L(λ) T, L(λ) T π(λ) = / 48

121 Future outlook Backward stability Backward stability Work ongoing in collaboration with many people (Piers Lawrence, Françoise Tisseur, Raf Vandebril, Paul Van Dooren,... ). I want to solve the eigenvalue problem L(λ) via QZ, with [ ] W T L(λ) = L(λ) T, L(λ) T π(λ) = but QZ will give me back the exact eigenvalues of L(λ): [ W L(λ) = T + δw T ] [ W = T ] L(λ) + δl(λ) L(λ) 42 / 48

122 Future outlook Backward stability Relating the two pencils The perturbation in W T is fine (it is a perturbation to the coefficients, so it seems reasonable.) 43 / 48

123 Future outlook Backward stability Relating the two pencils The perturbation in W T is fine (it is a perturbation to the coefficients, so it seems reasonable.) The perturbation to L(λ) will correspond to a perturbation to the dual space, that is I need to find δπ(λ) such that (L(λ) + δl(λ)) T (π(λ) + δπ(λ)) = / 48

124 Future outlook Backward stability Relating the two pencils The perturbation in W T is fine (it is a perturbation to the coefficients, so it seems reasonable.) The perturbation to L(λ) will correspond to a perturbation to the dual space, that is I need to find δπ(λ) such that (L(λ) + δl(λ)) T (π(λ) + δπ(λ)) = 0. Many interesting questions for us: When do small perturbations to L(λ) correspond to small perturbations on π(λ)? 43 / 48

125 Future outlook Backward stability Relating the two pencils The perturbation in W T is fine (it is a perturbation to the coefficients, so it seems reasonable.) The perturbation to L(λ) will correspond to a perturbation to the dual space, that is I need to find δπ(λ) such that (L(λ) + δl(λ)) T (π(λ) + δπ(λ)) = 0. Many interesting questions for us: When do small perturbations to L(λ) correspond to small perturbations on π(λ)? Can I measure the growth factor in the perturbation size? 43 / 48

126 Future outlook Backward stability Relating the two pencils The perturbation in W T is fine (it is a perturbation to the coefficients, so it seems reasonable.) The perturbation to L(λ) will correspond to a perturbation to the dual space, that is I need to find δπ(λ) such that (L(λ) + δl(λ)) T (π(λ) + δπ(λ)) = 0. Many interesting questions for us: When do small perturbations to L(λ) correspond to small perturbations on π(λ)? Can I measure the growth factor in the perturbation size? Does it work flawlessly also when using two bases (or more?). 43 / 48

127 Future outlook Backward stability To be continued to know the answer to all the above questions, you will need to see the next episode of dual bases and matrix polynomials. 44 / 48

128 Future outlook Backward stability To be continued to know the answer to all the above questions, you will need to see the next episode of dual bases and matrix polynomials. Likely the talk of Piers Lawrence at ILAS 2016 will be about these matters. 44 / 48

129 Future outlook Backward stability To be continued to know the answer to all the above questions, you will need to see the next episode of dual bases and matrix polynomials. Likely the talk of Piers Lawrence at ILAS 2016 will be about these matters. I have a draft in my laptop which seems promising. I am working on it in these days, let me know if you want to discuss something about that! 44 / 48

130 Future outlook Multivariate linearizations Multivariate polynomials What we have seen also naturally fit into the framework of linearizing the bivariate polynomial p(λ, µ) = n n p ij λ i µ j i=0 j=0 45 / 48

131 Future outlook Multivariate linearizations Multivariate polynomials What we have seen also naturally fit into the framework of linearizing the bivariate polynomial p(λ, µ) = n n p ij λ i µ j i=0 j=0 In fact, this can be naturally expressed as p(λ, µ) = π(λ) T Pπ(µ), P = (p ij ), and π(λ) the usual vector with the monomials. 45 / 48

132 Future outlook Multivariate linearizations Visualizing the linearization The following is a linearization for p(λ, µ): [ ] P L(λ) L(λ, µ) = L(µ) T 0 in the sense that det L(λ, µ) = p(λ, µ). 46 / 48

133 Future outlook Multivariate linearizations Visualizing the linearization The following is a linearization for p(λ, µ): [ ] P L(λ) L(λ, µ) = L(µ) T 0 in the sense that det L(λ, µ) = p(λ, µ). One could use this to solve problems of the form { p(λ, µ) = 0 q(λ, µ) = 0 by turning them into a multiparameter eigenvalue problem { L 1 (λ, µ)v 1 = 0 L 2 (λ, µ)v 2 = 0 46 / 48

134 Future outlook Multivariate linearizations Issues The previous step is not free from numerical issues. To solve the MEP one usually turns it into a linear eigenvalue problem, which makes the dimension grow, and add a quite large singular part to the pencil. 47 / 48

135 Future outlook Multivariate linearizations Issues The previous step is not free from numerical issues. To solve the MEP one usually turns it into a linear eigenvalue problem, which makes the dimension grow, and add a quite large singular part to the pencil. Work on this topic is being carried out by Bor Plestenjak. Our framework easily allows extension to more variables, which seems to be almost optimal in dimension for generic problems. 47 / 48

136 Conclusions Conclusions We have seen as dual bases and matrix polynomials play well together. 48 / 48

137 Conclusions Conclusions We have seen as dual bases and matrix polynomials play well together. This approach gives us a lot of flexibility in building linearizations, and gives us a better understanding of how different bases work. Moreover, we can combine different bases in the construction. 48 / 48

138 Conclusions Conclusions We have seen as dual bases and matrix polynomials play well together. This approach gives us a lot of flexibility in building linearizations, and gives us a better understanding of how different bases work. Moreover, we can combine different bases in the construction. Easier analysis of backward stability issues. Results are promising and seem to give new insights on scalings and linearization constructions. 48 / 48

139 Conclusions Conclusions We have seen as dual bases and matrix polynomials play well together. This approach gives us a lot of flexibility in building linearizations, and gives us a better understanding of how different bases work. Moreover, we can combine different bases in the construction. Easier analysis of backward stability issues. Results are promising and seem to give new insights on scalings and linearization constructions. Promising developments for multivariate rootfinding. 48 / 48

140 Conclusions Conclusions We have seen as dual bases and matrix polynomials play well together. This approach gives us a lot of flexibility in building linearizations, and gives us a better understanding of how different bases work. Moreover, we can combine different bases in the construction. Easier analysis of backward stability issues. Results are promising and seem to give new insights on scalings and linearization constructions. Promising developments for multivariate rootfinding. Thanks for your attention! 48 / 48