Page 191 - Matrices theory and applications
P. 191

10. Approximation of Eigenvalues
                              174
                              A j+1 := R j Q j .Wethenhave
                                                               −1
                                                      A j+1 = Q
                                                                 A j Q j ,
                                                               j
                              which shows that A j+1 is unitarily similar to A j . Hence,
                                                               −1
                                                                 A(Q 0 ··· Q j−1 )
                                              A j =(Q 0 ··· Q j−1 )
                              is conjugate to A by a unitary transformation.             (10.2)
                                Let P j := Q 0 ··· Q j−1 , which is unitary. Since U n is compact, the se-
                              quence (P j ) j∈IN possesses cluster values. Let P be one of them. Then
                              A := P −1 AP = P AP is a cluster point of (A j ) j∈IN . Hence, if the se-

                                               ∗
                              quence (A j ) j converges, its limit is unitarily similar to A, hence has the
                              same spectrum.
                                This argument shows that in general, the sequence (A j ) j does not con-
                              verge to a diagonal matrix, because then the eigenvectors of A would be
                              the columns of P.Inother words, A would have an orthonormal eigenba-
                              sis. Namely, A would be normal. Except in this special case, one expects
                              merely that the sequence (A j ) j converges to a triangular matrix, an expec-
                              tation that is compatible with Theorem 3.1.3. But even this hope is too
                              optimistic in general. For example, if A is unitary, then A j = A for every j,
                              with Q j = A and R j = I n ; in that case, the convergence is useless, since the
                              limit A is not simpler than the data. We shall see later on that the reason
                              for this bad behavior is that the eigenvalues of a unitary matrix have the
                              same modulus: The QR method does not do a good job of separating the
                              eigenvalues of close modulus.
                                An important case in which a matrix has at least two eigenvalues of
                              the same modulus is that of matrices with real entries. If A ∈ M n (IR),
                              then each Q j is real orthogonal, R j is real, and A j is real. This is seen by
                              induction on j. A limit A will not be triangular if some eigenvalues of A

                              are nonreal, that is, if A possesses a pair of complex conjugate eigenvalues.
                                Let us sum up what can be expected in a brave new world. If all the
                              eigenvalues of A ∈ M n (CC) have distinct moduli, the sequence (A j ) j might
                              converge to a triangular matrix, or at least its lower triangular part might
                              converge to
                                                                     
                                                      λ 1
                                                      0
                                                         λ 2         
                                                    .    .    .       .
                                                   
                                                                      
                                                    . .   . .  .  .  
                                                      0   ···  0   λ n
                              When A ∈ M n (IR), one makes the following assumption. Let p be the
                              number of real eigenvalues and 2q that of nonreal eigenvalues; then there
                              are p + q distinct eigenvalue moduli. In that case, (A j ) j might converge to
                              a block-triangular form, the diagonal blocks being 2×2or1×1. The limits
                              of the diagonal blocks provide trivially the eigenvalues of A.
   186   187   188   189   190   191   192   193   194   195   196