Schur s Triangularization Theorem. Math 422

Transcription

1 Schur s Triangularization Theorem Math 4 The characteristic polynomial p (t) of a square complex matrix A splits as a product of linear factors of the form (t λ) m Of course, finding these factors is a difficult problem, but having factored p (t) we can triangularize A whether or not A is diagonalizable Example 1 The characteristic polynomial p (t) t of the triangular matrix 0 1 A 0 0 hasthesinglerootλ 0, which is an eigenvalue of algebraic multiplicity The eigenspace of λ is one dimensional and is spanned by the single vector 1 0, so the geometric multiplicity of λ is 1 Therefore A is defective and is not diagonalizable (one needs three linearly independent eigenvectors to construct a transition matrix P that diagonalizes A) Let V n be an n-dimensional complex inner product space with Euclidean inner product Definition A hyperplane in V n is a translation of an (n 1)-dimensional subspace Note that the orthogonal complement u of a non-zero vector u C n is a hyperplane through the origin Consider the matrix P I 1 kuk uu ; then Q P 1 kuk uu I kuk uu is the Householder matrix associated with u Proposition N (P ) span {u} and multiplication by P is orthogonal projection on u, ie, for all x C n, Px proj u x Proof If Px 0, then x x u x u u 0or equivalently x u Thus x tu for some t C, and kuk kuk N (P )span {u} Furthermore, for all x C n, Px x proj u x proj u x Definition 4 Let x, y, u C n with u 6 0 Then y is the reflection of x in the hyperplane u iff x y proj u x Proposition 5 Let u C n be a non-zero vector reflection in the hyperplane u The Householder transformation associated with u is Proof Note that for all x C n, Thus Qx x Ã! x u kuk u x proj u x x Qx proj u x 1

2 Exercise 6 Earlier we proved that a real Householder matrix Q is symmetric and orthogonal, ie,q T Q and Q 1 Q T Generalize this result for complex matrices: Prove that complex Householder matrices Q are Hermitian and unitary Let v (v 1,,v n ) beanon-zerovectorinc n and set x v 1v kv 1 vk Then x (x 1,,x n ) C n is a unit vector with x 1 R Letu x e 1 then kuk (x e 1 ) (x e 1 )x x x e 1 e 1 x + e 1 e 1 x 1 (1 x e 1 ) x u x (x e 1 )x x x e 1 1 x e 1 If x 6 e 1, then u 6 0and we may apply the Householder transformation Q associated with u to x and e 1 : Qx x x u kuk u x (1 x e 1) (1 x e 1 ) u x u x (x e 1)e 1 ; applying Q to both sides we have Q x x Qe 1 If x e 1, set Q I; then in either case x Qe 1 and e 1 Qx are reflection of each other in the hyperplane (x e 1 ) For a unit vector x R, the line (x e 1 ) bisects the angle between x and e 1 We are ready to prove our main theorem in this lecture: Theorem 7 (Schur s Triangularization Theorem) Every square complex matrix A is unitarily similar to an upper-triangular matrix, ie, there exists a unitary matrix U such that T U AU is triangular Proof Use induction on the size of A For n 1there is nothing to prove So assume n>1 and that the result holds for all matrices of size less than n Since every complex matrix has an eigenvalue, choose an eigenvalue λ of A and an associated eigenvector v (v 1,,v n ) Let x v 1v kv 1 vk and set u x e 1; if x 6 e 1, let Q be the Householder matrix associated with u; if x e 1 let Q I Then x Qe 1 by the discussion above, so x is the first column of Q ByExercise6,Q is Hermitian and unitary, so x is the first row of Q Since Q Q 1 Q we have Q [x V ] x V and QAQ QA [x V ]Q [λx AV ] λe 1 x V λ AV x AV 0 V AV Now apply the induction hypothesis to V AV, which is an (n 1) (n 1) matrix, and obtain an (n 1) (n 1) unitary matrix R such that T n 1 R (V AV ) R is upper-triangular Let 1 0 U Q ; then U 1 0 U Q Q I R so U is unitary Hence T U AU QAQ 1 0 λ x AV V AV 1 0 λ x AV R 0 V AV R λ x AV R λ x V AV R AV R 0 T n 1

3 is triangular as claimed

4 Remark 8 Since similar matrices have the same eigenvalues, the eigenvalues of A are the diagonal entries of every Schur triangularization T U AU When all eigenvalues of A are real, Schur s Triangularization Theorem tells us that A is orthogonally similar to a triangular matrix Our next example demonstrates this Example 9 Let s find a Schur triangularization of the matrix 1 1 A The eigenvalues of A are λ 1 1,λ and λ Arbitrarily choose an eigenvalue, say λ 1 1, then 1 A I and x 1/ / is an associated unit eigenvector Let u x e 1 4/ / and let Q be the associated / / Householder matrix, ie, Q I 4 uut [x V ], 1 where V Then QAQ and V T AV Now triangularize the matrix V T AV, which has the single eigenvalue The vector x is a unit vector associated with Let u x e and let R be the Householder matrix associated with u, ie, R Then RV T 17/ AV R 0 is a Schur triangularization of V T AV Finally,let U Q 1 0 then U T AU is a (numerically approximate) Schur triangularization of A 4

5 Exercise 10 Show that the matrix A in Example 9 is defective and hence not diagonalizable In summary, every matrix is triangularizable but only non-defective matrices are diagonalizable Following the proof of Schur s Triangularization Theorem, find an orthogonal matrix P such that P T AP is upper triancular: 1 1 Exercise 11 A 1 Exercise 1 A Exercise 1 A