Outline. Scientific Computing: An Introductory Survey. Eigenvalue Problems. Eigenvalues and Eigenvectors. Geometric Interpretation

Transcription

1 Outline Scientific Computing: An Introductory Survey Chapter 4 Prof. Michael T. Heath Department of Computer Science University of Illinois at Urbana-Champaign Copyright c Reproduction permitted for noncommercial, educational use only. 2 3 Michael T. Heath Scientific Computing / 87 Michael T. Heath Scientific Computing 2 / 87 Eigenvalues and Eigenvectors Eigenvalues and Eigenvectors Eigenvalues and Eigenvectors Eigenvalue problems occur in many areas of science and engineering, such as structural analysis Eigenvalues are also important in analyzing numerical methods Theory and algorithms apply to complex matrices as well as real matrices With complex matrices, we use conjugate transpose, A H, instead of usual transpose, A T Standard eigenvalue problem : Given n n matrix A, find scalar λ and nonzero vector x such that A x λ x λ is eigenvalue, and x is corresponding eigenvector λ may be complex even if A is real Spectrum λ(a) set of eigenvalues of A Spectral radius ρ(a) max{ λ : λ λ(a)} Michael T. Heath Scientific Computing 3 / 87 Michael T. Heath Scientific Computing 4 / 87 Eigenvalues and Eigenvectors Eigenvalues and Eigenvectors Examples: Eigenvalues and Eigenvectors Matrix expands or shrinks any vector lying in direction of eigenvector by scalar factor Expansion or contraction factor is given by corresponding eigenvalue λ Eigenvalues and eigenvectors decompose complicated behavior of general linear transformation into simpler actions 0 0 A : λ 0 2, x, λ 0 2 2, x 2 A : λ 0 2, x, λ 0 2 2, x 2 3 A : λ 3 2, x, λ 2 4, x A : λ , x, λ 2, x 2 0 i A : λ 0 i, x, λ i 2 i, x 2 where i Michael T. Heath Scientific Computing 5 / 87 Michael T. Heath Scientific Computing 6 / 87 Example: Equation Ax λx is equivalent to (A λi)x 0 which has nonzero solution x if, and only if, its matrix is singular Eigenvalues of A are roots λ i of characteristic polynomial det(a λi) 0 in λ of degree n Fundamental Theorem of Algebra implies that n n matrix A always has n eigenvalues, but they may not be real nor distinct Complex eigenvalues of real matrix occur in complex conjugate pairs: if α + iβ is eigenvalue of real matrix, then so is α iβ, where i Michael T. Heath Scientific Computing 7 / 87 Characteristic polynomial of previous example matrix is ( ) 3 0 det λ 3 0 () 3 λ det 3 λ (3 λ)(3 λ) ( )( ) λ 2 6λ so eigenvalues are given by λ 6 ± 36 32, or λ 2, λ Michael T. Heath Scientific Computing 8 / 87

2 Companion Matrix, continued Monic polynomial p(λ) c 0 + c λ + + c n λ n + λ n is characteristic polynomial of companion matrix c c C n 0 0 c c n Roots of polynomial of degree > 4 cannot always computed in finite number of steps So in general, computation of eigenvalues of matrices of order > 4 requires (theoretically infinite) iterative process Michael T. Heath Scientific Computing 9 / 87 Example: Computing eigenvalues using characteristic polynomial is not recommended because of work in computing coefficients of characteristic polynomial sensitivity of coefficients of characteristic polynomial work in solving for roots of characteristic polynomial Characteristic polynomial is powerful theoretical tool but usually not useful computationally Multiplicity and Diagonalizability Michael T. Heath Scientific Computing 0 / 87 Consider A ɛ ɛ where ɛ is positive number slightly smaller than ɛ mach Exact eigenvalues of A are + ɛ and ɛ Computing characteristic polynomial in floating-point arithmetic, we obtain det(a λi) λ 2 2λ + ( ɛ 2 ) λ 2 2λ + which has as double root Thus, eigenvalues cannot be resolved by this method even though they are distinct in working precision Michael T. Heath Scientific Computing / 87 Eigenspaces and Invariant Subspaces Eigenvectors can be scaled arbitrarily: if Ax λx, then A(γx) λ(γx) for any scalar γ, so γx is also eigenvector corresponding to λ Eigenvectors are usually normalized by requiring some norm of eigenvector to be Eigenspace S λ {x : Ax λx} Multiplicity is number of times root appears when polynomial is written as product of linear factors Eigenvalue of multiplicity is simple Defective matrix has eigenvalue of multiplicity k > with fewer than k linearly independent corresponding eigenvectors Nondefective matrix A has n linearly independent eigenvectors, so it is diagonalizable X AX D where X is nonsingular matrix of eigenvectors Michael T. Heath Scientific Computing 2 / 87 Properties of matrix A relevant to eigenvalue problems Property Definition diagonal a ij 0 for i j tridiagonal a ij 0 for i j > triangular a ij 0 for i > j (upper) a ij 0 for i < j (lower) Hessenberg a ij 0 for i > j + (upper) a ij 0 for i < j (lower) Subspace S of R n (or C n ) is invariant if AS S For eigenvectors x x p, span([x x p ]) is invariant subspace orthogonal unitary symmetric Hermitian normal A T A AA T I A H A AA H I A A T A A H A H A AA H Examples: Matrix Properties Michael T. Heath Scientific Computing 3 / 87 T 2 3 Transpose: [ + i + 2i Conjugate transpose: 2 i 2 2i 2 Symmetric: Nonsymmetric: i Hermitian: NonHermitian: i 2 + i + i 2 ] H [ i ] 2 + i 2i 2 + 2i Examples, continued Michael T. Heath Scientific Computing 4 / Orthogonal:,, 0 0 [ ] i 2/2 2/2 Unitary: 2/2 i 2/2 Nonorthogonal: Normal: Nonnormal: 0 [ ] 2/2 2/2 2/2 2/2 Michael T. Heath Scientific Computing 5 / 87 Michael T. Heath Scientific Computing 6 / 87

3 Properties of of Properties of eigenvalue problem affecting choice of algorithm and software Are all eigenvalues needed, or only a few? Are only eigenvalues needed, or are corresponding eigenvectors also needed? Is matrix real or complex? Is matrix relatively small and dense, or large and sparse? Condition of eigenvalue problem is sensitivity of eigenvalues and eigenvectors to changes in matrix of eigenvalue problem is not same as conditioning of solution to linear system for same matrix Different eigenvalues and eigenvectors are not necessarily equally sensitive to perturbations in matrix Does matrix have any special properties, such as symmetry, or is it general matrix? of Eigenvalues Michael T. Heath Scientific Computing 7 / 87 of Eigenvalues Michael T. Heath Scientific Computing 8 / 87 If µ is eigenvalue of perturbation A + E of nondefective matrix A, then µ λ k cond 2 (X) E 2 where λ k is closest eigenvalue of A to µ and X is nonsingular matrix of eigenvectors of A Absolute condition number of eigenvalues is condition number of matrix of eigenvectors with respect to solving linear equations Eigenvalues may be sensitive if eigenvectors are nearly linearly dependent (i.e., matrix is nearly defective) For normal matrix (A H A AA H ), eigenvectors are orthogonal, so eigenvalues are well-conditioned Michael T. Heath Scientific Computing 9 / 87 If (A + E)(x + x) (λ + λ)(x + x), where λ is simple eigenvalue of A, then λ y 2 x 2 y H E 2 x cos(θ) E 2 where x and y are corresponding right and left eigenvectors and θ is angle between them For symmetric or Hermitian matrix, right and left eigenvectors are same, so cos(θ) and eigenvalues are inherently well-conditioned Eigenvalues of nonnormal matrices may be sensitive For multiple or closely clustered eigenvalues, corresponding eigenvectors may be sensitive Michael T. Heath Scientific Computing 20 / 87 Similarity Transformation Shift : If Ax λx and σ is any scalar, then (A σi)x (λ σ)x, so eigenvalues of shifted matrix are shifted eigenvalues of original matrix Inversion : If A is nonsingular and Ax λx with x 0, then λ 0 and A x (/λ)x, so eigenvalues of inverse are reciprocals of eigenvalues of original matrix Powers : If Ax λx, then A k x λ k x, so eigenvalues of power of matrix are same power of eigenvalues of original matrix Polynomial : If Ax λx and p(t) is polynomial, then p(a)x p(λ)x, so eigenvalues of polynomial in matrix are values of polynomial evaluated at eigenvalues of original matrix Michael T. Heath Scientific Computing 2 / 87 B is similar to A if there is nonsingular matrix T such that Then B T A T By λy T AT y λy A(T y) λ(t y) so A and B have same eigenvalues, and if y is eigenvector of B, then x T y is eigenvector of A Similarity transformations preserve eigenvalues and eigenvectors are easily recovered Michael T. Heath Scientific Computing 22 / 87 Example: Similarity Transformation Diagonal Form From eigenvalues and eigenvectors for previous example, and hence So original matrix is similar to diagonal matrix, and eigenvectors form columns of similarity transformation matrix Eigenvalues of diagonal matrix are diagonal entries, and eigenvectors are columns of identity matrix Diagonal form is desirable in simplifying eigenvalue problems for general matrices by similarity transformations But not all matrices are diagonalizable by similarity transformation Closest one can get, in general, is Jordan form, which is nearly diagonal but may have some nonzero entries on first superdiagonal, corresponding to one or more multiple eigenvalues Michael T. Heath Scientific Computing 23 / 87 Michael T. Heath Scientific Computing 24 / 87

4 Triangular Form Block Triangular Form Any matrix can be transformed into triangular (Schur ) form by similarity, and eigenvalues of triangular matrix are diagonal entries Eigenvectors of triangular matrix less obvious, but still straightforward to compute If U u U 3 A λi 0 0 v T O 0 U 33 is triangular, then U y u can be solved for y, so that y x 0 is corresponding eigenvector Forms Attainable by Similarity Michael T. Heath Scientific Computing 25 / 87 A T B distinct eigenvalues nonsingular diagonal real symmetric orthogonal real diagonal complex Hermitian unitary real diagonal normal unitary diagonal arbitrary real orthogonal real block triangular (real Schur) arbitrary unitary upper triangular (Schur) arbitrary nonsingular almost diagonal (Jordan) Given matrix A with indicated property, matrices B and T exist with indicated properties such that B T AT If B is diagonal or triangular, eigenvalues are its diagonal entries If B is diagonal, eigenvectors are columns of T Convergence of Power Iteration Michael T. Heath Scientific Computing 27 / 87 To see why power iteration converges to dominant eigenvector, express starting vector x 0 as linear combination n x 0 α i v i i where v i are eigenvectors of A Then x k Ax k A 2 x k 2 A k x 0 ( ) n n λ k i α i v i λ k α v + (λ i /λ ) k α i v i i Since λ i /λ < for i >, successively higher powers go to zero, leaving only component corresponding to v Limitations of Power Iteration i2 Michael T. Heath Scientific Computing 29 / 87 If A A 2 A p A 22 A 2p A.... with square diagonal blocks, then p λ(a) λ(a jj ) j A pp so eigenvalue problem breaks into p smaller eigenvalue problems Real Schur form has diagonal blocks corresponding to real eigenvalues and 2 2 diagonal blocks corresponding to pairs of complex conjugate eigenvalues Power Iteration Michael T. Heath Scientific Computing 26 / 87 Simplest method for computing one eigenvalueeigenvector pair is power iteration, which repeatedly multiplies matrix times initial starting vector Assume A has unique eigenvalue of maximum modulus, say λ, with corresponding eigenvector v Then, starting from nonzero vector x 0, iteration scheme x k Ax k converges to multiple of eigenvector v corresponding to dominant eigenvalue λ Example: Power Iteration Michael T. Heath Scientific Computing 28 / 87 Ratio of values of given component of x k from one iteration to next converges to [ dominant ] eigenvalue [ λ ] For example, if A and x , we obtain k x T k ratio Ratio is converging to dominant eigenvalue, which is 2 Normalized Power Iteration Michael T. Heath Scientific Computing 30 / 87 Power iteration can fail for various reasons Starting vector may have no component in dominant eigenvector v (i.e., α 0) not problem in practice because rounding error usually introduces such component in any case There may be more than one eigenvalue having same (maximum) modulus, in which case iteration may converge to linear combination of corresponding eigenvectors For real matrix and starting vector, iteration can never converge to complex vector Geometric growth of components at each iteration risks eventual overflow (or underflow if λ < ) Approximate eigenvector should be normalized at each iteration, say, by requiring its largest component to be in modulus, giving iteration scheme y k Ax k x k y k / y k With normalization, y k λ, and x k v / v Michael T. Heath Scientific Computing 3 / 87 Michael T. Heath Scientific Computing 32 / 87

5 Example: Normalized Power Iteration Repeating previous example with normalized scheme, k x T k y k < interactive example > Michael T. Heath Scientific Computing 33 / 87 Behavior of power iteration depicted geometrically Initial vector x 0 v + v 2 contains equal components in eigenvectors v and v 2 (dashed arrows) Repeated multiplication by A causes component in v (corresponding to larger eigenvalue, 2) to dominate, so sequence of vectors x k converges to v Michael T. Heath Scientific Computing 34 / 87 Power Iteration with Shift Example: Power Iteration with Shift Convergence rate of power iteration depends on ratio λ 2 /λ, where λ 2 is eigenvalue having second largest modulus May be possible to choose shift, A σi, such that λ 2 σ λ σ < λ 2 so convergence is accelerated λ In earlier example, for instance, if we pick shift of σ, (which is equal to other eigenvalue) then ratio becomes zero and method converges in one iteration In general, we would not be able to make such fortuitous choice, but shifts can still be extremely useful in some contexts, as we will see later Shift must then be added to result to obtain eigenvalue of original matrix Michael T. Heath Scientific Computing 35 / 87 Michael T. Heath Scientific Computing 36 / 87 Inverse Iteration Inverse Iteration, continued If smallest eigenvalue of matrix required rather than largest, can make use of fact that eigenvalues of A are reciprocals of those of A, so smallest eigenvalue of A is reciprocal of largest eigenvalue of A This leads to inverse iteration scheme Ay k x k x k y k / y k Inverse iteration converges to eigenvector corresponding to smallest eigenvalue of A Eigenvalue obtained is dominant eigenvalue of A, and hence its reciprocal is smallest eigenvalue of A in modulus which is equivalent to power iteration applied to A Inverse of A not computed explicitly, but factorization of A used to solve system of linear equations at each iteration Michael T. Heath Scientific Computing 37 / 87 Michael T. Heath Scientific Computing 38 / 87 Example: Inverse Iteration Inverse Iteration with Shift Applying inverse iteration to previous example to compute smallest eigenvalue yields sequence k x T k y k which is indeed converging to (which is its own reciprocal in this case) < interactive example > As before, shifting strategy, working with A σi for some scalar σ, can greatly improve convergence Inverse iteration is particularly useful for computing eigenvector corresponding to approximate eigenvalue, since it converges rapidly when applied to shifted matrix A λi, where λ is approximate eigenvalue Inverse iteration is also useful for computing eigenvalue closest to given value β, since if β is used as shift, then desired eigenvalue corresponds to smallest eigenvalue of shifted matrix Michael T. Heath Scientific Computing 39 / 87 Michael T. Heath Scientific Computing 40 / 87

6 Rayleigh Quotient Example: Rayleigh Quotient Given approximate eigenvector x for real matrix A, determining best estimate for corresponding eigenvalue λ can be considered as n linear least squares approximation problem xλ Ax From normal equation x T xλ x T Ax, least squares solution is given by λ xt Ax x T x This quantity, known as Rayleigh quotient, has many useful properties Rayleigh Quotient Iteration Michael T. Heath Scientific Computing 4 / 87 Rayleigh quotient can accelerate convergence of iterative methods such as power iteration, since Rayleigh quotient x T k Ax k/x T k x k gives better approximation to eigenvalue at iteration k than does basic method alone For previous example using power iteration, value of Rayleigh quotient at each iteration is shown below k x T k y k x T k Ax k/x T k x k Michael T. Heath Scientific Computing 42 / 87 Rayleigh Quotient Iteration, continued Given approximate eigenvector, Rayleigh quotient yields good estimate for corresponding eigenvalue Conversely, inverse iteration converges rapidly to eigenvector if approximate eigenvalue is used as shift, with one iteration often sufficing These two ideas combined in Rayleigh quotient iteration σ k x T k Ax k/x T k x k (A σ k I)y k+ x k x k+ y k+ / y k+ Rayleigh quotient iteration is especially effective for symmetric matrices and usually converges very rapidly Using different shift at each iteration means matrix must be refactored each time to solve linear system, so cost per iteration is high unless matrix has special form that makes factorization easy Same idea also works for complex matrices, for which transpose is replaced by conjugate transpose, so Rayleigh quotient becomes x H Ax/x H x starting from given nonzero vector x 0 Michael T. Heath Scientific Computing 43 / 87 Michael T. Heath Scientific Computing 44 / 87 Example: Rayleigh Quotient Iteration Deflation Using same matrix as previous examples and randomly chosen starting vector x 0, Rayleigh quotient iteration converges in two iterations k x T k σ k Michael T. Heath Scientific Computing 45 / 87 After eigenvalue λ and corresponding eigenvector x have been computed, then additional eigenvalues λ 2,..., λ n of A can be computed by deflation, which effectively removes known eigenvalue Let H be any nonsingular matrix such that Hx αe, scalar multiple of first column of identity matrix (Householder transformation is good choice for H) Then similarity transformation determined by H transforms A into form [ HAH λ b T ] 0 B where B is matrix of order n having eigenvalues λ 2,..., λ n Michael T. Heath Scientific Computing 46 / 87 Deflation, continued Deflation, continued Thus, we can work with B to compute next eigenvalue λ 2 Moreover, if y 2 is eigenvector of B corresponding to λ 2, then x 2 H [ α y 2 ], where α bt y 2 λ 2 λ is eigenvector corresponding to λ 2 for original matrix A, provided λ λ 2 Process can be repeated to find additional eigenvalues and eigenvectors Alternative approach lets u be any vector such that u T x λ Then A x u T has eigenvalues 0, λ 2,..., λ n Possible choices for u include u λ x, if A is symmetric and x is normalized so that x 2 u λ y, where y is corresponding left eigenvector (i.e., A T y λ y ) normalized so that y T x u A T e k, if x is normalized so that x and kth component of x is Michael T. Heath Scientific Computing 47 / 87 Michael T. Heath Scientific Computing 48 / 87

7 Simultaneous Iteration Orthogonal Iteration Simplest method for computing many eigenvalueeigenvector pairs is simultaneous iteration, which repeatedly multiplies matrix times matrix of initial starting vectors Starting from n p matrix X 0 of rank p, iteration scheme is X k AX k span(x k ) converges to invariant subspace determined by p largest eigenvalues of A, provided λ p > λ p+ Also called subspace iteration Michael T. Heath Scientific Computing 49 / 87 As with power iteration, normalization is needed with simultaneous iteration Each column of X k converges to dominant eigenvector, so columns of X k become increasingly ill-conditioned basis for span(x k ) Both issues can be addressed by computing QR factorization at each iteration ˆQ k R k X k X k A ˆQ k where ˆQ k R k is reduced QR factorization of X k This orthogonal iteration converges to block triangular form, and leading block is triangular if moduli of consecutive eigenvalues are distinct Michael T. Heath Scientific Computing 50 / 87 QR Iteration QR Iteration, continued For p n and X 0 I, matrices A k ˆQ H k A ˆQ k generated by orthogonal iteration converge to triangular or block triangular form, yielding all eigenvalues of A QR iteration computes successive matrices A k without forming above product explicitly Starting with A 0 A, at iteration k compute QR factorization Q k R k A k and form reverse product A k R k Q k Successive matrices A k are unitarily similar to each other A k R k Q k Q H k A k Q k Diagonal entries (or eigenvalues of diagonal blocks) of A k converge to eigenvalues of A Product of orthogonal matrices Q k converges to matrix of corresponding eigenvectors If A is symmetric, then symmetry is preserved by QR iteration, so A k converge to matrix that is both triangular and symmetric, hence diagonal Example: QR Iteration Michael T. Heath Scientific Computing 5 / Let A Compute QR factorization A 0 Q R and form reverse product A R Q [ 7.83 ] Off-diagonal entries are now smaller, and diagonal entries closer to eigenvalues, 8 and 3 Process continues until matrix is within tolerance of being diagonal, and diagonal entries then closely approximate eigenvalues Michael T. Heath Scientific Computing 53 / 87 QR Iteration with Shifts Michael T. Heath Scientific Computing 52 / 87 Convergence rate of QR iteration can be accelerated by incorporating shifts Q k R k A k σ k I A k R k Q k + σ k I where σ k is rough approximation to eigenvalue Good shift can be determined by computing eigenvalues of 2 2 submatrix in lower right corner of matrix Michael T. Heath Scientific Computing 54 / 87 Example: QR Iteration with Shifts Preliminary Reduction Repeat previous example, but with shift of σ 4, which is lower right corner entry of matrix We compute QR factorization A 0 σ I Q R and form reverse product, adding back shift to obtain A R Q + σ I After one iteration, off-diagonal entries smaller compared with unshifted algorithm, and eigenvalues closer approximations to eigenvalues < interactive example > Efficiency of QR iteration can be enhanced by first transforming matrix as close to triangular form as possible before beginning iterations Hessenberg matrix is triangular except for one additional nonzero diagonal immediately adjacent to main diagonal Any matrix can be reduced to Hessenberg form in finite number of steps by orthogonal similarity transformation, for example using Householder transformations Symmetric Hessenberg matrix is tridiagonal Hessenberg or tridiagonal form is preserved during successive QR iterations Michael T. Heath Scientific Computing 55 / 87 Michael T. Heath Scientific Computing 56 / 87

8 Preliminary Reduction, continued Advantages of initial reduction to upper Hessenberg or tridiagonal form Work per QR iteration is reduced from O(n 3 ) to O(n 2 ) for general matrix or O(n) for symmetric matrix Fewer QR iterations are required because matrix nearly triangular (or diagonal) already If any zero entries on first subdiagonal, then matrix is block triangular and problem can be broken into two or more smaller subproblems Michael T. Heath Scientific Computing 57 / 87 Preliminary Reduction, continued QR iteration is implemented in two-stages symmetric tridiagonal diagonal or general Hessenberg triangular Preliminary reduction requires definite number of steps, whereas subsequent iterative stage continues until convergence In practice only modest number of iterations usually required, so much of work is in preliminary reduction Cost of accumulating eigenvectors, if needed, dominates total cost Michael T. Heath Scientific Computing 58 / 87 Cost of QR Iteration Krylov Subspace Methods Approximate overall cost of preliminary reduction and QR iteration, counting both additions and multiplications Symmetric matrices 4 3 n3 for eigenvalues only 9n 3 for eigenvalues and eigenvectors General matrices 0n 3 for eigenvalues only 25n 3 for eigenvalues and eigenvectors Michael T. Heath Scientific Computing 59 / 87 Krylov subspace methods reduce matrix to Hessenberg (or tridiagonal) form using only matrix-vector multiplication For arbitrary starting vector x 0, if K k [ x 0 Ax 0 A k x 0 ] then Kn AK n C n where C n is upper Hessenberg (in fact, companion matrix) To obtain better conditioned basis for span(k n ), compute QR factorization Q n R n K n so that Q H n AQ n R n C n Rn H with H upper Hessenberg Michael T. Heath Scientific Computing 60 / 87 Krylov Subspace Methods Arnoldi Iteration Equating kth columns on each side of equation AQ n Q n H yields recurrence Aq k h k q + + h kk q k + h k+,k q k+ relating q k+ to preceding vectors q,..., q k Premultiplying by q H j and using orthonormality, h jk q H j Aq k, j,..., k These relationships yield Arnoldi iteration, which produces unitary matrix Q n and upper Hessenberg matrix H n column by column using only matrix-vector multiplication by A and inner products of vectors x 0 arbitrary nonzero starting vector q x 0 / x 0 2 for k, 2,... u k Aq k for j to k h jk q H j u k u k u k h jk q j end h k+,k u k 2 if h k+,k 0 then stop q k+ u k /h k+,k end Arnoldi Iteration, continued Michael T. Heath Scientific Computing 6 / 87 Arnoldi Iteration, continued Michael T. Heath Scientific Computing 62 / 87 If then is upper Hessenberg matrix Q k [ q q k ] H k Q H k AQ k Eigenvalues of H k, called Ritz values, are approximate eigenvalues of A, and Ritz vectors given by Q k y, where y is eigenvector of H k, are corresponding approximate eigenvectors of A Eigenvalues of H k must be computed by another method, such as QR iteration, but this is easier problem if k n Arnoldi iteration fairly expensive in work and storage because each new vector q k must be orthogonalized against all previous columns of Q k, and all must be stored for that purpose. So Arnoldi process usually restarted periodically with carefully chosen starting vector Ritz values and vectors produced are often good approximations to eigenvalues and eigenvectors of A after relatively few iterations Michael T. Heath Scientific Computing 63 / 87 Michael T. Heath Scientific Computing 64 / 87

9 Lanczos Iteration Work and storage costs drop dramatically if matrix is symmetric or Hermitian, since recurrence then has only three terms and H k is tridiagonal (so usually denoted T k ) q 0 0 β 0 0 x 0 arbitrary nonzero starting vector q x 0 / x 0 2 for k, 2,... u k Aq k α k q H k u k u k u k β k q k α k q k β k u k 2 if β k 0 then stop q k+ u k /β k end Lanczos Iteration, continued Michael T. Heath Scientific Computing 65 / 87 Lanczos Iteration, continued α k and β k are diagonal and subdiagonal entries of symmetric tridiagonal matrix T k As with Arnoldi, Lanczos iteration does not produce eigenvalues and eigenvectors directly, but only tridiagonal matrix T k, whose eigenvalues and eigenvectors must be computed by another method to obtain Ritz values and vectors If β k 0, then algorithm appears to break down, but in that case invariant subspace has already been identified (i.e., Ritz values and vectors are already exact at that point) Michael T. Heath Scientific Computing 66 / 87 Krylov Subspace Methods, continued In principle, if Lanczos algorithm were run until k n, resulting tridiagonal matrix would be orthogonally similar to A In practice, rounding error causes loss of orthogonality, invalidating this expectation Problem can be overcome by reorthogonalizing vectors as needed, but expense can be substantial Alternatively, can ignore problem, in which case algorithm still produces good eigenvalue approximations, but multiple copies of some eigenvalues may be generated Great virtue of Arnoldi and Lanczos iterations is their ability to produce good approximations to extreme eigenvalues for k n Moreover, they require only one matrix-vector multiplication by A per step and little auxiliary storage, so are ideally suited to large sparse matrices If eigenvalues are needed in middle of spectrum, say near σ, then algorithm can be applied to matrix (A σi), assuming it is practical to solve systems of form (A σi)x y Example: Lanczos Iteration Michael T. Heath Scientific Computing 67 / 87 For symmetric matrix with eigenvalues,..., 29, behavior of Lanczos iteration is shown below Michael T. Heath Scientific Computing 69 / 87 Jacobi Method Michael T. Heath Scientific Computing 68 / 87 One of oldest methods for computing eigenvalues is Jacobi method, which uses similarity transformation based on plane rotations Sequence of plane rotations chosen to annihilate symmetric pairs of matrix entries, eventually converging to diagonal form Choice of plane rotation slightly more complicated than in Givens method for QR factorization To annihilate given off-diagonal pair, choose c and s so that J T c s a b c s AJ s c b d s c [ c 2 a 2csb + s 2 d c 2 b + cs(a d) s 2 ] b c 2 b + cs(a d) s 2 b c 2 d + 2csb + s 2 a is diagonal Michael T. Heath Scientific Computing 70 / 87 Jacobi Method, continued Example: Plane Rotation Transformed matrix diagonal if c 2 b + cs(a d) s 2 b 0 Dividing both sides by c 2 b, we obtain + s (a d) s2 c b c 2 0 Making substitution t s/c, we get quadratic equation (a d) + t t 2 0 b for tangent t of angle of rotation, from which we can recover c /( + t 2 ) and s c t Advantageous numerically to use root of smaller magnitude Michael T. Heath Scientific Computing 7 / 87 Consider 2 2 matrix A 2 2 Quadratic equation for tangent reduces to t 2, so t ± Two roots of same magnitude, so we arbitrarily choose t, which yields c / 2 and s / 2 Resulting plane rotation J gives [ ] [ ] J T / 2 / 2 AJ / 2 / 2 / 2 / / 2 / Michael T. Heath Scientific Computing 72 / 87

10 Jacobi Method, continued Jacobi Method, continued Starting with symmetric matrix A 0 A, each iteration has form A k+ J T k A kj k where J k is plane rotation chosen to annihilate a symmetric pair of entries in A k Plane rotations repeatedly applied from both sides in systematic sweeps through matrix until off-diagonal mass of matrix is reduced to within some tolerance of zero Resulting diagonal matrix orthogonally similar to original matrix, so diagonal entries are eigenvalues, and eigenvectors are given by product of plane rotations Jacobi method is reliable, simple to program, and capable of high accuracy, but converges rather slowly and difficult to generalize beyond symmetric matrices Except for small problems, more modern methods usually require 5 to 0 times less work than Jacobi One source of inefficiency is that previously annihilated entries can subsequently become nonzero again, thereby requiring repeated annihilation Newer methods such as QR iteration preserve zero entries introduced into matrix Example: Jacobi Method 0 2 Let A Michael T. Heath Scientific Computing 73 / 87 First annihilate (,3) and (3,) entries using rotation J to obtain A J0 T A 0 J Example, continued Michael T. Heath Scientific Computing 74 / 87 Next annihilate (,2) and (2,) entries using rotation J to obtain A 2 J T A J Michael T. Heath Scientific Computing 75 / 87 Michael T. Heath Scientific Computing 76 / 87 Example, continued Example, continued Next annihilate (2,3) and (3,2) entries using rotation 0 0 J to obtain A 3 J2 T A 2 J Beginning new sweep, again annihilate (,3) and (3,) entries using rotation J to obtain A 4 J3 T A 3 J Michael T. Heath Scientific Computing 77 / 87 Michael T. Heath Scientific Computing 78 / 87 Example, continued Bisection or Spectrum-Slicing Process continues until off-diagonal entries reduced to as small as desired Result is diagonal matrix orthogonally similar to original matrix, with the orthogonal similarity transformation given by product of plane rotations < interactive example > For real symmetric matrix, can determine how many eigenvalues are less than given real number σ By systematically choosing various values for σ (slicing spectrum at σ) and monitoring resulting count, any eigenvalue can be isolated as accurately as desired For example, symmetric indefinite factorization A LDL T makes inertia (numbers of positive, negative, and zero eigenvalues) of symmetric matrix A easy to determine By applying factorization to matrix A σi for various values of σ, individual eigenvalues can be isolated as accurately as desired using interval bisection technique Michael T. Heath Scientific Computing 79 / 87 Michael T. Heath Scientific Computing 80 / 87

11 Sturm Sequence Divide-and-Conquer Method Another spectrum-slicing method for computing individual eigenvalues is based on Sturm sequence property of symmetric matrices Let A be symmetric matrix and let p r (σ) denote determinant of leading principal minor of order r of A σi Then zeros of p r (σ) strictly separate those of p r (σ), and number of agreements in sign of successive members of sequence p r (σ), for r,..., n, equals number of eigenvalues of A strictly greater than σ Determinants p r (σ) are easy to compute if A is transformed to tridiagonal form before applying Sturm sequence technique Express symmetric tridiagonal matrix T as T O T + β uu T O T 2 Can now compute eigenvalues and eigenvectors of smaller matrices T and T 2 To relate these back to eigenvalues and eigenvectors of original matrix requires solution of secular equation, which can be done reliably and efficiently Applying this approach recursively yields divide-and-conquer algorithm for symmetric tridiagonal eigenproblems Relatively Robust Representation Michael T. Heath Scientific Computing 8 / 87 With conventional methods, cost of computing eigenvalues of symmetric tridiagonal matrix is O(n 2 ), but if orthogonal eigenvectors are also computed, then cost rises to O(n 3 ) Another possibility is to compute eigenvalues first at O(n 2 ) cost, and then compute corresponding eigenvectors separately using inverse iteration with computed eigenvalues as shifts Key to making this idea work is computing eigenvalues and corresponding eigenvectors to very high relative accuracy so that expensive explicit orthogonalization of eigenvectors is not needed RRR algorithm exploits this approach to produce eigenvalues and orthogonal eigenvectors at O(n 2 ) cost Michael T. Heath Scientific Computing 83 / 87 Michael T. Heath Scientific Computing 82 / 87 Generalized Generalized eigenvalue problem has form Ax λbx where A and B are given n n matrices If either A or B is nonsingular, then generalized eigenvalue problem can be converted to standard eigenvalue problem, either (B A)x λx or (A B)x (/λ)x This is not recommended, since it may cause loss of accuracy due to rounding error loss of symmetry if A and B are symmetric Better alternative for generalized eigenvalue problems is QZ algorithm Michael T. Heath Scientific Computing 84 / 87 QZ Algorithm QZ Algorithm, continued If A and B are triangular, then eigenvalues are given by λ i a ii /b ii, for b ii 0 QZ algorithm reduces A and B simultaneously to upper triangular form by orthogonal transformations First, B is reduced to upper triangular form by orthogonal transformation from left, which is also applied to A Next, transformed A is reduced to upper Hessenberg form by orthogonal transformation from left, while maintaining triangular form of B, which requires additional transformations from right Finally, analogous to QR iteration, A is reduced to triangular form while still maintaining triangular form of B, which again requires transformations from both sides Eigenvalues can now be determined from mutually triangular form, and eigenvectors can be recovered from products of left and right transformations, denoted by Q and Z Michael T. Heath Scientific Computing 85 / 87 Michael T. Heath Scientific Computing 86 / 87 Computing SVD Singular values of A are nonnegative square roots of eigenvalues of A T A, and columns of U and V are orthonormal eigenvectors of AA T and A T A, respectively Algorithms for computing SVD work directly with A, without forming AA T or A T A, to avoid loss of information associated with forming these matrix products explicitly SVD is usually computed by variant of QR iteration, with A first reduced to bidiagonal form by orthogonal transformations, then remaining off-diagonal entries are annihilated iteratively SVD can also be computed by variant of Jacobi method, which can be useful on parallel computers or if matrix has special structure Michael T. Heath Scientific Computing 87 / 87