Home Random quiz Cards

Chapter 6

Eigenvalue Problems

Rendered from the uploaded full LaTeX source with KaTeX. No PDF, no PDF.js, no image extraction, and no raw LaTeX displayed.

Learning goals

After completing this chapter, the reader should be able to:

  1. distinguish algebraic, geometric, and numerical aspects of eigenvalues;

  2. use Gershgorin disks to localize eigenvalues;

  3. derive the power method and its convergence rate;

  4. analyze inverse iteration, shifted inverse iteration, and Rayleigh quotient iteration;

  5. understand Rayleigh quotients, residuals, and a posteriori eigenvalue error estimates;

  6. prove basic eigenvalue perturbation results;

  7. derive orthogonal iteration and the QR algorithm;

  8. understand Hessenberg reduction, shifted QR, and symmetric tridiagonal QR;

  9. formulate Arnoldi, Lanczos, Ritz, and Krylov eigenvalue approximations;

  10. understand preconditioning, pseudospectra, generalized eigenvalue problems, and large-scale eigenvalue computations;

  11. solve basic, advanced, and research-level exercises on eigenvalue problems.

The Eigenvalue Problem

The standard eigenvalue problem is Ax=\lambda x, \qquad x\ne0, where A\in\mathbb C^{n\times n}, \lambda\in\mathbb C, and x\in\mathbb C^n. The scalar \lambda is an eigenvalue, and x is an associated eigenvector.

Equivalently, (A-\lambda I)x=0 has a nontrivial solution. Thus \det(A-\lambda I)=0. The polynomial p_A(\lambda)=\det(\lambda I-A) is the characteristic polynomial of A.

Key point: Central idea

Eigenvalue algorithms are not primarily about expanding the characteristic polynomial. Modern methods transform matrices while preserving eigenvalues, or approximate selected eigenvalues using Krylov subspaces.

Definition: Spectrum and spectral radius

The spectrum of A is \sigma(A)=\{\lambda\in\mathbb C:\det(\lambda I-A)=0\}. The spectral radius is \rho(A)=\max_{\lambda\in\sigma(A)}|\lambda|.

Theorem: Existence of eigenvalues over C

Every matrix A\in\mathbb C^{n\times n} has at least one eigenvalue.

Proof

The characteristic polynomial p_A(\lambda)=\det(\lambda I-A) has degree n. By the fundamental theorem of algebra, it has at least one complex root. If p_A(\lambda)=0, then \lambda I-A is singular, so there exists a nonzero vector x such that (\lambda I-A)x=0. Hence Ax=\lambda x.

Algebraic and Geometric Multiplicity

The algebraic multiplicity of an eigenvalue \lambda is its multiplicity as a root of the characteristic polynomial. The geometric multiplicity is \dim\ker(A-\lambda I).

Theorem: Geometric multiplicity is bounded by algebraic multiplicity

For every eigenvalue \lambda, 1\le \dim\ker(A-\lambda I) \le \text{algebraic multiplicity of }\lambda.

Proof

Let g=\dim\ker(A-\lambda I). Choose a basis of the eigenspace and extend it to a basis of \mathbb C^n. In this basis, A has block form \begin{pmatrix} \lambda I_g & *\\ 0 & B \end{pmatrix}. Therefore \det(tI-A) = (t-\lambda)^g\det(tI-B). Thus (t-\lambda)^g divides the characteristic polynomial, so the algebraic multiplicity is at least g.

Diagonalization and Schur Decomposition

A matrix A is diagonalizable if there exists a nonsingular matrix X such that A=X\Lambda X^{-1}, where \Lambda=\operatorname{diag}(\lambda_1,\ldots,\lambda_n).

Theorem: Diagonalization criterion

A matrix A\in\mathbb C^{n\times n} is diagonalizable if and only if it has n linearly independent eigenvectors.

Proof

If A=X\Lambda X^{-1}, then AX=X\Lambda. Thus the columns of X are eigenvectors of A, and they are linearly independent because X is nonsingular.

Conversely, if A has n linearly independent eigenvectors x_1,\ldots,x_n, let X=[x_1\ \cdots\ x_n], \qquad \Lambda=\operatorname{diag}(\lambda_1,\ldots,\lambda_n). Then AX=X\Lambda. Multiplying by X^{-1} gives A=X\Lambda X^{-1}.

Theorem: Schur decomposition

For every A\in\mathbb C^{n\times n}, there exists a unitary matrix Q and an upper triangular matrix T such that A=QTQ^*. The diagonal entries of T are the eigenvalues of A.

Proof

We argue by induction on n. By the referenced result, A has an eigenpair (\lambda,q_1) with \|q_1\|_2=1. Extend q_1 to an orthonormal basis and let Q_1=[q_1\ Q_2]. Then Q_1^*AQ_1 = \begin{pmatrix} \lambda & *\\ 0 & A_2 \end{pmatrix}. By the induction hypothesis, A_2=Q_3T_2Q_3^*. Then Q=Q_1 \begin{pmatrix} 1&0\\ 0&Q_3 \end{pmatrix} unitarily reduces A to upper triangular form. The eigenvalues of an upper triangular matrix are its diagonal entries.

Chapter summary: Why Schur form matters

The Schur decomposition always exists and uses a unitary similarity transformation. This makes it fundamental in stable eigenvalue algorithms: unitary transformations preserve the 2-norm and avoid unnecessary conditioning loss.

Normal and Hermitian Matrices

A matrix is normal if A^*A=AA^*. Hermitian matrices, unitary matrices, and real symmetric matrices are normal.

Theorem: Spectral theorem for Hermitian matrices

If A=A^*, then all eigenvalues of A are real and there exists a unitary matrix Q such that A=Q\Lambda Q^*, where \Lambda is real diagonal.

Proof

If Ax=\lambda x, with x\ne0, then x^*Ax=\lambda x^*x. Since A=A^*, the scalar x^*Ax is real. Hence \lambda is real.

By the Schur decomposition, A=QTQ^*, where T is upper triangular. Since A is Hermitian, T is also Hermitian: T^*=T. An upper triangular Hermitian matrix must be diagonal. Therefore A=Q\Lambda Q^* with \Lambda real diagonal.

Theorem: Min-max principle

Let A=A^* have eigenvalues \lambda_1\le \lambda_2\le\cdots\le\lambda_n. Then \lambda_k = \min_{\dim S=k} \max_{\substack{x\in S\\x\ne0}} \frac{x^*Ax}{x^*x}.

Proof

Let q_1,\ldots,q_n be an orthonormal eigenbasis. For any k-dimensional subspace S, the intersection of S with \operatorname{span}\{q_k,\ldots,q_n\} is nontrivial. Therefore for some nonzero x\in S, \frac{x^*Ax}{x^*x}\ge\lambda_k. Thus the maximum over S is at least \lambda_k. Taking S=\operatorname{span}\{q_1,\ldots,q_k\} makes the maximum equal to \lambda_k. This proves the result.

Gershgorin Localization

Theorem: Gershgorin disk theorem

Every eigenvalue of A=(a_{ij})\in\mathbb C^{n\times n} lies in at least one disk D_i= \left\{ z\in\mathbb C: |z-a_{ii}|\le \sum_{j\ne i}|a_{ij}| \right\}.

Proof

Let Ax=\lambda x, and choose an index i such that |x_i|=\max_j|x_j|. Then x_i\ne0. The i-th component of Ax=\lambda x gives (\lambda-a_{ii})x_i=\sum_{j\ne i}a_{ij}x_j. Taking absolute values, |\lambda-a_{ii}|\,|x_i| \le \sum_{j\ne i}|a_{ij}|\,|x_j| \le \sum_{j\ne i}|a_{ij}|\,|x_i|. Dividing by |x_i| gives |\lambda-a_{ii}| \le \sum_{j\ne i}|a_{ij}|.

Figure 6.1 Eigenvector direction Eigenvectors are directions that are only scaled by the matrix.
Open visual gallery

Rayleigh Quotient

For A=A^*, the Rayleigh quotient is \rho(x)=\frac{x^*Ax}{x^*x}, \qquad x\ne0. If x is an eigenvector, then \rho(x)=\lambda.

Theorem: Rayleigh quotient residual orthogonality

Let \mu=\rho(x)=\frac{x^*Ax}{x^*x}. Then the residual r=Ax-\mu x is orthogonal to x: x^*r=0.

Proof

Compute x^*r = x^*Ax-\mu x^*x. By the definition of \mu, \mu x^*x=x^*Ax. Thus x^*r=0.

Theorem: Rayleigh quotient bounds for Hermitian matrices

Let A=A^*, with eigenvalues \lambda_{\min}\le\lambda_i\le\lambda_{\max}. Then for all x\ne0, \lambda_{\min} \le \rho(x) \le \lambda_{\max}.

Proof

Let A=Q\Lambda Q^*, and set y=Q^*x. Then \rho(x) = \frac{y^*\Lambda y}{y^*y} = \frac{\sum_i \lambda_i |y_i|^2}{\sum_i |y_i|^2}. This is a weighted average of the eigenvalues, with nonnegative weights. Hence it lies between the smallest and largest eigenvalues.

A Posteriori Eigenvalue Error Bounds

Suppose \|x\|_2=1, \qquad r=Ax-\mu x. If A is Hermitian, small residual implies that \mu is close to some eigenvalue.

Theorem: Residual eigenvalue bound for Hermitian matrices

Let A=A^*, \|x\|_2=1, and r=Ax-\mu x. Then there exists an eigenvalue \lambda\in\sigma(A) such that |\lambda-\mu|\le \|r\|_2.

Proof

Let A=Q\Lambda Q^*, and set y=Q^*x. Since \|x\|_2=1, \sum_i |y_i|^2=1. Also, \|r\|_2^2 = \|(\Lambda-\mu I)y\|_2^2 = \sum_i |\lambda_i-\mu|^2 |y_i|^2. If every eigenvalue satisfied |\lambda_i-\mu|>\|r\|_2, then the last sum would be greater than \|r\|_2^2\sum_i|y_i|^2=\|r\|_2^2, a contradiction. Hence at least one eigenvalue satisfies the stated bound.

Power Method

Assume A is diagonalizable: A=X\Lambda X^{-1}, with eigenvalues ordered so that |\lambda_1|>|\lambda_2|\ge\cdots\ge|\lambda_n|. Let x^{(0)}=\sum_{j=1}^n c_jv_j, \qquad c_1\ne0. Then A^kx^{(0)} = c_1\lambda_1^kv_1 + \sum_{j=2}^n c_j\lambda_j^kv_j. After normalization, the direction approaches v_1.

Algorithm
Caption.

Power method

  1. matrix A, initial vector x^{(0)}\ne0

  2. x^{(0)}\gets x^{(0)}/\|x^{(0)}\|_2

  3. For k=0,1,2,\ldots:

  4. y^{(k)}\gets Ax^{(k)}

  5. x^{(k+1)}\gets y^{(k)}/\|y^{(k)}\|_2

  6. \mu_{k+1}\gets (x^{(k+1)})^*Ax^{(k+1)}

  7. If \|Ax^{(k+1)}-\mu_{k+1}x^{(k+1)}\|_2 is small enough:

  8. Return \mu_{k+1},x^{(k+1)}

Theorem: Power method convergence

Assume A is diagonalizable and |\lambda_1|>|\lambda_2|\ge\cdots\ge|\lambda_n|. If the initial vector has a nonzero component in the direction of v_1, then the normalized power iterates converge in direction to v_1, with asymptotic factor \left|\frac{\lambda_2}{\lambda_1}\right|.

Proof

Write x^{(0)}=\sum_{j=1}^n c_jv_j, \qquad c_1\ne0. Then A^kx^{(0)} = \lambda_1^k \left[ c_1v_1+\sum_{j=2}^n c_j \left(\frac{\lambda_j}{\lambda_1}\right)^k v_j \right]. Since \left|\frac{\lambda_j}{\lambda_1}\right| \le \left|\frac{\lambda_2}{\lambda_1}\right|<1, the bracketed expression converges to c_1v_1. Normalization removes the scalar factor \lambda_1^k, so the direction converges to that of v_1.

Figure 6.2 Power iteration alignment Repeated multiplication aligns with the dominant eigenmode.
Open visual gallery

Inverse Iteration and Shifted Inverse Iteration

Inverse iteration applies the power method to A^{-1}. If |\lambda_j| is the smallest eigenvalue magnitude, then 1/\lambda_j is dominant for A^{-1}.

Shifted inverse iteration solves (A-\sigma I)y^{(k)}=x^{(k)}, \qquad x^{(k+1)}=\frac{y^{(k)}}{\|y^{(k)}\|}. The eigenvalues of (A-\sigma I)^{-1} are \frac{1}{\lambda_j-\sigma}. Thus the method converges toward the eigenvalue closest to the shift \sigma.

Algorithm
Caption.

Shifted inverse iteration

  1. matrix A, shift \sigma, initial vector x^{(0)}\ne0

  2. Factorize A-\sigma I

  3. x^{(0)}\gets x^{(0)}/\|x^{(0)}\|_2

  4. For k=0,1,2,\ldots:

  5. Solve (A-\sigma I)y^{(k)}=x^{(k)}

  6. x^{(k+1)}\gets y^{(k)}/\|y^{(k)}\|_2

  7. \mu_{k+1}\gets (x^{(k+1)})^*Ax^{(k+1)}

  8. If \|Ax^{(k+1)}-\mu_{k+1}x^{(k+1)}\|_2 is small enough:

  9. Return \mu_{k+1},x^{(k+1)}

Theorem: Shifted inverse iteration convergence factor

Suppose A is diagonalizable and \lambda_\ast is the eigenvalue closest to \sigma. If |\lambda_\ast-\sigma| < |\lambda_j-\sigma| \qquad j\ne\ast, then shifted inverse iteration converges to the eigenvector associated with \lambda_\ast, with asymptotic factor \max_{j\ne\ast} \left| \frac{\lambda_\ast-\sigma}{\lambda_j-\sigma} \right|.

Proof

Apply the power-method theorem to (A-\sigma I)^{-1}. Its eigenvalues are \theta_j=\frac{1}{\lambda_j-\sigma}. The dominant eigenvalue corresponds to the smallest |\lambda_j-\sigma|, namely \lambda_\ast. Therefore the power-method convergence factor is \max_{j\ne\ast} \left| \frac{\theta_j}{\theta_\ast} \right| = \max_{j\ne\ast} \left| \frac{\lambda_\ast-\sigma}{\lambda_j-\sigma} \right|.

Rayleigh Quotient Iteration

Rayleigh quotient iteration updates the shift dynamically: \mu_k=\rho(x^{(k)}), (A-\mu_kI)y^{(k)}=x^{(k)}, x^{(k+1)}=\frac{y^{(k)}}{\|y^{(k)}\|_2}.

Algorithm
Caption.

Rayleigh quotient iteration

  1. matrix A, initial vector x^{(0)}\ne0

  2. x^{(0)}\gets x^{(0)}/\|x^{(0)}\|_2

  3. For k=0,1,2,\ldots:

  4. \mu_k\gets (x^{(k)})^*Ax^{(k)}

  5. Solve (A-\mu_k I)y^{(k)}=x^{(k)}

  6. x^{(k+1)}\gets y^{(k)}/\|y^{(k)}\|_2

  7. If \|Ax^{(k+1)}-\mu_kx^{(k+1)}\|_2 is small enough:

  8. Return \mu_k,x^{(k+1)}

Theorem: Local cubic convergence of RQI for Hermitian matrices

Let A=A^*, and suppose RQI starts sufficiently close to a simple eigenvector. Then the eigenvector error converges cubically.

Proof

For Hermitian A, the Rayleigh quotient error is second order in the angle to the eigenvector. More precisely, if x is close to a normalized eigenvector q, then |\rho(x)-\lambda|=O(\sin^2\theta), where \theta is the angle between x and q. Shifted inverse iteration with shift \mu=\rho(x) has local factor proportional to |\lambda-\mu|=O(\sin^2\theta). Multiplying this factor by the current angular error gives \sin\theta_{\text{new}}=O(\sin^3\theta). Thus the convergence is cubic.

Orthogonal Iteration and Simultaneous Iteration

Power iteration approximates one dominant eigenvector. Orthogonal iteration approximates an invariant subspace. Starting with an orthonormal matrix Q_0, one computes Z_k=AQ_{k-1}, then forms the QR factorization Z_k=Q_kR_k.

Algorithm
Caption.

Orthogonal iteration

  1. matrix A, initial orthonormal matrix Q_0

  2. For k=1,2,\ldots:

  3. Z_k\gets AQ_{k-1}

  4. Compute QR factorization Z_k=Q_kR_k

  5. Return Q_k

The QR Algorithm

The basic QR algorithm starts with A_0=A. At step k, A_k=Q_kR_k, A_{k+1}=R_kQ_k. Since A_{k+1}=Q_k^*A_kQ_k, the matrices A_k are unitarily similar and therefore have the same eigenvalues.

Algorithm
Caption.

Basic QR algorithm

  1. matrix A

  2. A_0\gets A

  3. For k=0,1,2,\ldots:

  4. Compute A_k=Q_kR_k

  5. A_{k+1}\gets R_kQ_k

  6. Return diagonal or quasi-triangular form of A_k

Theorem: QR iteration preserves eigenvalues

The matrices generated by the QR algorithm satisfy A_{k+1}=Q_k^*A_kQ_k. Therefore all A_k are unitarily similar and have the same eigenvalues.

Proof

Since A_k=Q_kR_k, we have R_k=Q_k^*A_k. Thus A_{k+1}=R_kQ_k=Q_k^*A_kQ_k. Unitary similarity preserves the characteristic polynomial and therefore preserves the eigenvalues.

Hessenberg Reduction

Before applying QR iteration to a dense nonsymmetric matrix, one reduces it to upper Hessenberg form: h_{ij}=0 \qquad \text{for } i>j+1. This is done by Householder transformations: H=Q^*AQ. QR iteration preserves Hessenberg form and costs O(n^2) per iteration instead of O(n^3).

Algorithm
Caption.

Householder reduction to Hessenberg form

  1. A\in\mathbb C^{n\times n}

  2. For k=1,\ldots,n-2:

  3. Choose Householder reflector H_k to zero entries k+2,\ldots,n in column k

  4. A\gets H_k^*AH_k

  5. Return upper Hessenberg matrix H

Figure 6.3 Shifted inverse iteration The shift selects which eigenvalue is targeted.
Open visual gallery

Shifted QR Algorithm

A shifted QR step uses a shift \mu_k: A_k-\mu_kI=Q_kR_k, A_{k+1}=R_kQ_k+\mu_kI. Then A_{k+1}=Q_k^*A_kQ_k. Good shifts accelerate convergence dramatically. For symmetric tridiagonal matrices, the Wilkinson shift is a standard robust choice.

Chapter summary: Modern dense eigenvalue computation

A practical dense nonsymmetric eigenvalue solver typically follows this pipeline: A \longrightarrow \text{Hessenberg form} \longrightarrow \text{shifted QR iteration} \longrightarrow \text{Schur form}. For real symmetric matrices: A \longrightarrow \text{tridiagonal form} \longrightarrow \text{symmetric shifted QR}.

Jacobi Method for Symmetric Eigenvalue Problems

The Jacobi eigenvalue method applies plane rotations to a symmetric matrix to reduce the off-diagonal norm. For A=A^T, choose a pair (p,q) and apply a rotation J so that (J^TAJ)_{pq}=0. Repeated rotations drive the matrix toward diagonal form.

Algorithm
Caption.

Jacobi eigenvalue method

  1. symmetric matrix A

  2. While off-diagonal norm is not small:

  3. Choose an off-diagonal pivot a_{pq}

  4. Construct a plane rotation J that zeros the (p,q) and (q,p) entries

  5. A\gets J^TAJ

  6. Return approximate eigenvalues from diagonal entries

Arnoldi Method for Large Nonsymmetric Problems

For large sparse matrices, forming a dense Schur decomposition is impossible. Arnoldi constructs an orthonormal basis of \mathcal K_k(A,v)=\operatorname{span}\{v,Av,\ldots,A^{k-1}v\} and the relation AV_k=V_kH_k+h_{k+1,k}v_{k+1}e_k^T. The eigenvalues of H_k are Ritz values approximating eigenvalues of A.

Algorithm
Caption.

Arnoldi eigenvalue iteration

  1. matrix A, unit starting vector v_1, dimension k

  2. For j=1,\ldots,k:

  3. w\gets Av_j

  4. For i=1,\ldots,j:

  5. h_{ij}\gets v_i^*w

  6. w\gets w-h_{ij}v_i

  7. h_{j+1,j}\gets \|w\|_2

  8. If h_{j+1,j}=0:

  9. stop

  10. v_{j+1}\gets w/h_{j+1,j}

  11. Compute eigenpairs of H_k

  12. Return Ritz values and Ritz vectors V_ky

Theorem: Ritz residual formula for Arnoldi

Let H_ky=\theta y, \qquad \|y\|_2=1, and let u=V_ky. Then Au-\theta u = h_{k+1,k}v_{k+1}e_k^Ty, and therefore \|Au-\theta u\|_2 = |h_{k+1,k}|\,|e_k^Ty|.

Proof

Using the Arnoldi relation AV_k=V_kH_k+h_{k+1,k}v_{k+1}e_k^T, multiply by y: AV_ky = V_kH_ky+h_{k+1,k}v_{k+1}e_k^Ty. Since H_ky=\theta y, Au = \theta u+h_{k+1,k}v_{k+1}e_k^Ty. Taking norms and using \|v_{k+1}\|_2=1 gives the formula.

Lanczos Method for Hermitian Problems

For A=A^*, Arnoldi reduces to Lanczos: AV_k=V_kT_k+\beta_kv_{k+1}e_k^T, where T_k is real symmetric tridiagonal. The eigenvalues of T_k approximate eigenvalues of A.

Algorithm
Caption.

Lanczos iteration

  1. Hermitian A, unit starting vector v_1

  2. \beta_0\gets0, v_0\gets0

  3. For j=1,\ldots,k:

  4. w\gets Av_j-\beta_{j-1}v_{j-1}

  5. \alpha_j\gets v_j^*w

  6. w\gets w-\alpha_jv_j

  7. \beta_j\gets \|w\|_2

  8. If \beta_j=0:

  9. stop

  10. v_{j+1}\gets w/\beta_j

  11. Form tridiagonal T_k from \alpha_j and \beta_j

  12. Return Ritz values from T_k

Eigenvalue Conditioning

Let A have a simple eigenvalue \lambda, right eigenvector x, and left eigenvector y: Ax=\lambda x, \qquad y^*A=\lambda y^*. Normalize so that \|x\|_2=\|y\|_2=1. The eigenvalue condition number is \kappa(\lambda)=\frac{1}{|y^*x|}.

Theorem: First-order eigenvalue perturbation

Let \lambda be a simple eigenvalue of A, with right and left eigenvectors x and y. For a small perturbation E, \lambda(A+E) = \lambda + \frac{y^*Ex}{y^*x} + O(\|E\|^2). Consequently, |\Delta\lambda| \le \kappa(\lambda)\|E\|_2+O(\|E\|_2^2), when \|x\|_2=\|y\|_2=1.

Proof

Let (A+E)(x+\Delta x)=(\lambda+\Delta\lambda)(x+\Delta x). Keeping only first-order terms gives A\Delta x+Ex=\lambda\Delta x+\Delta\lambda x. Rearrange: (A-\lambda I)\Delta x+Ex=\Delta\lambda x. Left-multiply by y^*. Since y^*(A-\lambda I)=0, we obtain y^*Ex=\Delta\lambda\, y^*x. Thus \Delta\lambda=\frac{y^*Ex}{y^*x} to first order.

Warning: Non-normal matrices can be dangerous

For normal matrices, eigenvectors are orthogonal and eigenvalues are well behaved under perturbations. For non-normal matrices, eigenvalues can be extremely sensitive, and residuals may be misleading unless conditioning is considered.

Bauer–Fike Theorem and Pseudospectra

Theorem: Bauer–Fike theorem

Suppose A=X\Lambda X^{-1} is diagonalizable. If \mu is an eigenvalue of A+E, then there exists \lambda\in\sigma(A) such that |\mu-\lambda| \le \kappa_2(X)\|E\|_2.

Proof

If \mu\notin\sigma(A), then A-\mu I is invertible. If \mu is an eigenvalue of A+E, then A+E-\mu I is singular, so I+(A-\mu I)^{-1}E is singular. Hence \|(A-\mu I)^{-1}E\|_2\ge1. Thus \|(A-\mu I)^{-1}\|_2\|E\|_2\ge1. Since A-\mu I=X(\Lambda-\mu I)X^{-1}, we have \|(A-\mu I)^{-1}\|_2 \le \kappa_2(X) \max_{\lambda\in\sigma(A)}\frac{1}{|\lambda-\mu|}. Therefore \min_{\lambda\in\sigma(A)}|\lambda-\mu| \le \kappa_2(X)\|E\|_2.

The \varepsilon-pseudospectrum is \sigma_\varepsilon(A) = \{z\in\mathbb C:\|(zI-A)^{-1}\|>1/\varepsilon\}. It is equivalently the set of eigenvalues of all perturbed matrices A+E with \|E\|<\varepsilon.

Generalized Eigenvalue Problems

The generalized eigenvalue problem is Ax=\lambda Bx, \qquad x\ne0. If B is nonsingular, this can be written as B^{-1}Ax=\lambda x. However, explicitly forming B^{-1}A may be unstable or structure-destroying. The QZ algorithm generalizes the QR algorithm and computes generalized Schur form for the pencil A-\lambda B.

Singular Values as Eigenvalues

Singular values are eigenvalue-related quantities. The singular values of A\in\mathbb C^{m\times n} are the square roots of the eigenvalues of A^*A. However, forming A^*A squares the condition number and may lose accuracy. Stable SVD algorithms avoid explicitly forming the normal matrix when high accuracy is needed.

Practical Comparison of Eigenvalue Methods

Chapter summary: Choosing an eigenvalue method
Power method.

Simple and cheap. It finds the dominant eigenvalue when a clear spectral gap exists.

Inverse iteration.

Efficient for finding an eigenvalue near a chosen shift, provided linear solves with A-\sigma I are available.

Rayleigh quotient iteration.

Very fast locally for Hermitian problems, often cubic, but requires solving a shifted linear system at every step.

Dense QR algorithm.

The standard method for computing all eigenvalues of a dense matrix after Hessenberg or tridiagonal reduction.

Jacobi eigenvalue method.

Reliable for symmetric matrices and highly parallel in some forms, but usually not the fastest dense method.

Arnoldi.

The main Krylov method for large sparse nonsymmetric eigenvalue problems. It approximates selected eigenvalues through Ritz values.

Lanczos.

Efficient for large sparse Hermitian eigenvalue problems. It uses short recurrences but needs care with loss of orthogonality.

Shift-and-invert Krylov.

Powerful for interior eigenvalues. Its cost is dominated by solving shifted linear systems.

QZ algorithm.

The standard dense method for generalized eigenvalue problems Ax=\lambda Bx.

Exercises

The following exercise bank is intentionally large. Basic problems test definitions, computations, and essential facts. Starred exercises require proofs, careful derivations, stability analysis, or advanced matrix theory. Problems marked \star, \star\star, and \star\star\star are progressively harder.

Basic problems

Exercise 6.1 Basic Eigenpair verification

Verify whether x=\begin{pmatrix}1\\1\end{pmatrix} is an eigenvector of A=\begin{pmatrix}2&1\\1&2\end{pmatrix}. If it is, find the eigenvalue.

Exercise 6.2 Basic Characteristic polynomial

Compute the characteristic polynomial of A=\begin{pmatrix}1&2\\3&4\end{pmatrix}.

Exercise 6.3 Basic Eigenvalues of a triangular matrix

Explain why the eigenvalues of an upper triangular matrix are its diagonal entries.

Exercise 6.4 Basic Algebraic and geometric multiplicity

For A=\begin{pmatrix}1&1\\0&1\end{pmatrix}, compute the algebraic and geometric multiplicities of the eigenvalue 1.

Exercise 6.5 Basic Diagonalization

Diagonalize A=\begin{pmatrix}2&0\\0&3\end{pmatrix}.

Exercise 6.6 Basic Symmetric matrix eigenvalues

Explain why real symmetric matrices have real eigenvalues.

Exercise 6.7 Basic Rayleigh quotient

Compute the Rayleigh quotient of A=\begin{pmatrix}2&0\\0&5\end{pmatrix}, \qquad x=\begin{pmatrix}1\\1\end{pmatrix}.

Exercise 6.8 Basic Gershgorin disks

Find the Gershgorin disks of A=\begin{pmatrix} 4&1&0\\ 2&3&1\\ 0&1&2 \end{pmatrix}.

Exercise 6.9 Basic One power iteration

Apply one power iteration to A=\begin{pmatrix}2&1\\1&2\end{pmatrix}, \qquad x^{(0)}=\begin{pmatrix}1\\0\end{pmatrix}. Normalize using the Euclidean norm.

Exercise 6.10 Basic Dominant eigenvalue

For a diagonal matrix A=\operatorname{diag}(5,2,1), which eigenvalue is targeted by the power method for a generic initial vector?

Exercise 6.11 Basic Inverse iteration

Explain why inverse iteration targets eigenvalues close to the shift \sigma.

Exercise 6.12 Basic Shifted matrix

If A has eigenvalues 1,3,7, what are the eigenvalues of A-2I?

Exercise 6.13 Basic Residual norm

Given (\mu,x), define the eigenvalue residual and explain why it is useful.

Exercise 6.14 Basic QR similarity

Show that if A=QR, then RQ=Q^*AQ.

Exercise 6.15 Basic Hessenberg form

What entries of an upper Hessenberg matrix are necessarily zero?

Exercise 6.16 Basic Lanczos versus Arnoldi

What matrix property allows Arnoldi to reduce to the Lanczos process?

Exercise 6.17 Basic Ritz value

Define a Ritz value in a subspace method.

Exercise 6.18 Basic Generalized eigenvalue problem

State the generalized eigenvalue problem.

Exercise 6.19 Basic Singular values

How are singular values related to eigenvalues of A^*A?

Exercise 6.20 Basic Normal matrix

Define a normal matrix and give two examples.

Intermediate problems \star

Exercise 6.21 Intermediate Existence theorem

Prove the referenced result.

Exercise 6.22 Intermediate Multiplicity inequality

Prove the referenced result.

Exercise 6.23 Intermediate Diagonalization criterion

Prove the referenced result.

Exercise 6.24 Intermediate Schur decomposition

Prove the referenced result.

Exercise 6.25 Intermediate Spectral theorem

Prove the referenced result.

Exercise 6.26 Intermediate Rayleigh quotient bounds

Prove the referenced result.

Exercise 6.27 Intermediate Rayleigh residual orthogonality

Prove the referenced result.

Exercise 6.28 Intermediate Gershgorin theorem

Prove the referenced result.

Exercise 6.29 Intermediate Disjoint Gershgorin disks

State and prove the refinement of Gershgorin’s theorem for connected components of the union of disks.

Exercise 6.30 Intermediate Power method convergence

Prove the referenced result.

Exercise 6.31 Intermediate Power method failure

Give examples where the power method fails or stagnates.

Exercise 6.32 Intermediate Shifted inverse iteration factor

Prove the referenced result.

Exercise 6.33 Intermediate Rayleigh quotient iteration

Derive Rayleigh quotient iteration from shifted inverse iteration.

Exercise 6.34 Intermediate QR eigenvalue preservation

Prove the referenced result.

Exercise 6.35 Intermediate Shifted QR similarity

Show that a shifted QR step also produces a unitarily similar matrix.

Exercise 6.36 Intermediate Hessenberg preservation

Show that QR iteration preserves upper Hessenberg form.

Exercise 6.37 Intermediate Jacobi rotation

Derive the plane rotation that zeros the (p,q) entry of a real symmetric matrix.

Exercise 6.38 Intermediate Arnoldi relation

Derive the Arnoldi relation AV_k=V_kH_k+h_{k+1,k}v_{k+1}e_k^T.

Exercise 6.39 Intermediate Ritz residual

Prove the referenced result.

Exercise 6.40 Intermediate Lanczos recurrence

Derive the Lanczos three-term recurrence from Arnoldi when A=A^*.

Advanced problems \star\star

Exercise 6.41 Advanced Min-max principle

Prove the referenced result in full detail.

Exercise 6.42 Advanced Courant–Fischer theorem

State and prove both min-max and max-min forms of the Courant–Fischer theorem.

Exercise 6.43 Advanced Interlacing theorem

Prove the Cauchy interlacing theorem for eigenvalues of Hermitian matrices.

Exercise 6.44 Advanced Weyl inequalities

State and prove Weyl’s eigenvalue perturbation inequality for Hermitian matrices.

Exercise 6.45 Advanced Davis–Kahan theorem

State the Davis–Kahan sin-theta theorem and explain its meaning for invariant subspace perturbation.

Exercise 6.46 Advanced Residual eigenvalue bound

Prove the referenced result.

Exercise 6.47 Advanced Eigenvector residual bound

For Hermitian A, derive an angle bound between an approximate eigenvector and an exact eigenspace using a spectral gap.

Exercise 6.48 Advanced First-order perturbation

Prove the referenced result.

Exercise 6.49 Advanced Bauer–Fike theorem

Prove the referenced result.

Exercise 6.50 Advanced Pseudospectrum equivalence

Prove the equivalence between pseudospectra defined by resolvent norm and by eigenvalues of perturbed matrices.

Exercise 6.51 Advanced Non-normal sensitivity

Construct a non-normal matrix whose eigenvalues are highly sensitive to tiny perturbations.

Exercise 6.52 Advanced Rayleigh quotient accuracy

Show that for Hermitian matrices, the Rayleigh quotient eigenvalue error is second order in the eigenvector angle.

Exercise 6.53 Advanced RQI cubic convergence

Prove the referenced result more rigorously.

Exercise 6.54 Advanced Orthogonal iteration

Analyze convergence of orthogonal iteration to a dominant invariant subspace.

Exercise 6.55 Advanced QR and orthogonal iteration

Show the relationship between QR iteration and orthogonal iteration.

Exercise 6.56 Advanced Hessenberg reduction cost

Derive the operation count for Householder reduction to Hessenberg form.

Exercise 6.57 Advanced Symmetric tridiagonalization

Derive Householder tridiagonalization for real symmetric matrices.

Exercise 6.58 Advanced Wilkinson shift

Derive the Wilkinson shift for a symmetric tridiagonal QR step.

Exercise 6.59 Advanced Deflation

Explain deflation in the QR algorithm and derive a practical deflation criterion.

Exercise 6.60 Advanced Francis QR step

Describe the implicit Francis double-shift QR step for real nonsymmetric matrices.

Exercise 6.61 Advanced Lanczos ghost eigenvalues

Explain how loss of orthogonality in Lanczos can produce ghost eigenvalues.

Exercise 6.62 Advanced Selective reorthogonalization

Describe selective reorthogonalization in Lanczos methods.

Exercise 6.63 Advanced Restarted Arnoldi

Explain the need for restarting in Arnoldi eigenvalue computations.

Exercise 6.64 Advanced Implicitly restarted Arnoldi

Study the implicitly restarted Arnoldi method and its relation to shifted QR.

Exercise 6.65 Advanced Shift-and-invert Arnoldi

Analyze shift-and-invert Arnoldi for computing interior eigenvalues.

Research-level problems \star\star\star

Exercise 6.66 Research-level Backward stability of QR algorithm

Study the backward stability of the shifted QR algorithm and explain the role of unitary transformations.

Exercise 6.67 Research-level Aggressive early deflation

Investigate aggressive early deflation in modern QR eigensolvers.

Exercise 6.68 Research-level Divide-and-conquer symmetric eigenproblem

Study divide-and-conquer methods for symmetric tridiagonal eigenvalue problems.

Exercise 6.69 Research-level MRRR algorithm

Study the MRRR algorithm for symmetric tridiagonal eigenproblems and discuss its accuracy tradeoffs.

Exercise 6.70 Research-level Jacobi high relative accuracy

Analyze why Jacobi-type methods can deliver high relative accuracy for certain symmetric eigenvalue problems.

Exercise 6.71 Research-level Polynomial filtering

Design a Chebyshev polynomial filter for separating eigenvalues in a target interval.

Exercise 6.72 Research-level Rational Krylov methods

Study rational Krylov subspaces and compare them with polynomial Krylov subspaces.

Exercise 6.73 Research-level Jacobi–Davidson method

Derive the Jacobi–Davidson correction equation and explain how it targets selected eigenpairs.

Exercise 6.74 Research-level LOBPCG

Study the locally optimal block preconditioned conjugate-gradient method for Hermitian eigenproblems.

Exercise 6.75 Research-level FEAST algorithm

Explain the FEAST algorithm as a contour-integration spectral projector method.

Exercise 6.76 Research-level Generalized Hermitian definite eigenproblem

Analyze the problem Ax=\lambda Bx, \qquad A=A^*, \qquad B=B^*>0. Derive a reduction to a standard Hermitian eigenvalue problem.

Exercise 6.77 Research-level QZ algorithm

Study the QZ algorithm for generalized eigenvalue problems and compare it with QR.

Exercise 6.78 Research-level Polynomial eigenvalue problems

Study polynomial eigenvalue problems and linearization strategies.

Exercise 6.79 Research-level Nonlinear eigenvalue problems

Investigate nonlinear eigenvalue problems T(\lambda)x=0 and discuss Newton-type methods for them.

Exercise 6.80 Research-level Pseudospectral portraits

Design a numerical method for plotting pseudospectra using resolvent norms.

Exercise 6.81 Research-level Eigenvalues of discretized PDEs

Analyze eigenvalue computations for large sparse matrices arising from finite difference or finite element discretizations of elliptic operators.

Exercise 6.82 Research-level Randomized eigensolvers

Study randomized subspace iteration for approximating dominant eigenspaces.

Exercise 6.83 Research-level Communication-avoiding eigensolvers

Investigate communication-avoiding variants of Arnoldi or Lanczos methods.

Exercise 6.84 Research-level Mixed precision eigensolvers

Analyze the use of mixed precision in eigenvalue computations, including residual correction and refinement.

Exercise 6.85 Research-level Certified eigenvalue enclosures

Develop interval or residual-based certified enclosures for Hermitian eigenvalues.

Exercise solutions

The solutions below are converted from the uploaded LaTeX solution file. Open an accordion to view the full solution.

This chapter gives detailed worked solutions for the exercises in Chapter 6. Each solution includes the problem formulation, the method, the mathematical derivation, the conclusion, and a diagnostic comment.

Exercise 6.1

Problem formulation.

Verify whether x=\begin{pmatrix}1\\1\end{pmatrix} is an eigenvector of A=\begin{pmatrix}2&1\\1&2\end{pmatrix}. If it is, find the eigenvalue.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.2

Problem formulation.

Compute the characteristic polynomial of A=\begin{pmatrix}1&2\\3&4\end{pmatrix}.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.3

Problem formulation.

Explain why the eigenvalues of an upper triangular matrix are its diagonal entries.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.4

Problem formulation.

For A=\begin{pmatrix}1&1\\0&1\end{pmatrix}, compute the algebraic and geometric multiplicities of the eigenvalue 1.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.5

Problem formulation.

Diagonalize A=\begin{pmatrix}2&0\\0&3\end{pmatrix}.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.6

Problem formulation.

Explain why real symmetric matrices have real eigenvalues.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.7

Problem formulation.

Compute the Rayleigh quotient of A=\begin{pmatrix}2&0\\0&5\end{pmatrix}, \qquad x=\begin{pmatrix}1\\1\end{pmatrix}.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

For A=A^T, the Rayleigh quotient is \rho(x)=\frac{x^TAx}{x^Tx}. If x is an eigenvector, Ax=\lambda x, then \rho(x)=\frac{x^T\lambda x}{x^Tx}=\lambda. The stationary points of \rho(x) on the unit sphere satisfy Ax=\lambda x, so eigenvectors are precisely stationary points of the Rayleigh quotient.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.8

Problem formulation.

Find the Gershgorin disks of A=\begin{pmatrix} 4&1&0\\ 2&3&1\\ 0&1&2 \end{pmatrix}.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Let Ax=\lambda x, and choose i such that |x_i|=\max_j|x_j|. Then (\lambda-a_{ii})x_i=\sum_{j\ne i}a_{ij}x_j. Taking absolute values, |\lambda-a_{ii}||x_i| \le \sum_{j\ne i}|a_{ij}||x_j| \le \left(\sum_{j\ne i}|a_{ij}|\right)|x_i|. Since x_i\ne0, |\lambda-a_{ii}| \le \sum_{j\ne i}|a_{ij}|. Thus every eigenvalue lies in a Gershgorin disk.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.9

Problem formulation.

Apply one power iteration to A=\begin{pmatrix}2&1\\1&2\end{pmatrix}, \qquad x^{(0)}=\begin{pmatrix}1\\0\end{pmatrix}. Normalize using the Euclidean norm.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Assume A has eigenpairs (\lambda_j,v_j) with |\lambda_1|>|\lambda_2|\ge\cdots. Write x_0=\sum_j c_jv_j, \qquad c_1\ne0. Then A^kx_0=\lambda_1^k \left[ c_1v_1+\sum_{j\ge2}c_j \left(\frac{\lambda_j}{\lambda_1}\right)^k v_j \right]. After normalization, the terms with j\ge2 vanish geometrically at rate \left|\frac{\lambda_2}{\lambda_1}\right|^k. Thus the power method converges to the dominant eigenvector.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.10

Problem formulation.

For a diagonal matrix A=\operatorname{diag}(5,2,1), which eigenvalue is targeted by the power method for a generic initial vector?

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Assume A has eigenpairs (\lambda_j,v_j) with |\lambda_1|>|\lambda_2|\ge\cdots. Write x_0=\sum_j c_jv_j, \qquad c_1\ne0. Then A^kx_0=\lambda_1^k \left[ c_1v_1+\sum_{j\ge2}c_j \left(\frac{\lambda_j}{\lambda_1}\right)^k v_j \right]. After normalization, the terms with j\ge2 vanish geometrically at rate \left|\frac{\lambda_2}{\lambda_1}\right|^k. Thus the power method converges to the dominant eigenvector.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.11

Problem formulation.

Explain why inverse iteration targets eigenvalues close to the shift \sigma.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.12

Problem formulation.

If A has eigenvalues 1,3,7, what are the eigenvalues of A-2I?

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.13

Problem formulation.

Given (\mu,x), define the eigenvalue residual and explain why it is useful.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.14

Problem formulation.

Show that if A=QR, then RQ=Q^*AQ.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.15

Problem formulation.

What entries of an upper Hessenberg matrix are necessarily zero?

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.16

Problem formulation.

What matrix property allows Arnoldi to reduce to the Lanczos process?

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.17

Problem formulation.

Define a Ritz value in a subspace method.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.18

Problem formulation.

State the generalized eigenvalue problem.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.19

Problem formulation.

How are singular values related to eigenvalues of A^*A?

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.20

Problem formulation.

Define a normal matrix and give two examples.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.21

Problem formulation.

Prove the referenced result.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.22

Problem formulation.

Prove the referenced result.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.23

Problem formulation.

Prove the referenced result.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.24

Problem formulation.

Prove the referenced result.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.25

Problem formulation.

Prove the referenced result.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.26

Problem formulation.

Prove the referenced result.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

For A=A^T, the Rayleigh quotient is \rho(x)=\frac{x^TAx}{x^Tx}. If x is an eigenvector, Ax=\lambda x, then \rho(x)=\frac{x^T\lambda x}{x^Tx}=\lambda. The stationary points of \rho(x) on the unit sphere satisfy Ax=\lambda x, so eigenvectors are precisely stationary points of the Rayleigh quotient.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.27

Problem formulation.

Prove the referenced result.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

For A=A^T, the Rayleigh quotient is \rho(x)=\frac{x^TAx}{x^Tx}. If x is an eigenvector, Ax=\lambda x, then \rho(x)=\frac{x^T\lambda x}{x^Tx}=\lambda. The stationary points of \rho(x) on the unit sphere satisfy Ax=\lambda x, so eigenvectors are precisely stationary points of the Rayleigh quotient.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.28

Problem formulation.

Prove the referenced result.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Let Ax=\lambda x, and choose i such that |x_i|=\max_j|x_j|. Then (\lambda-a_{ii})x_i=\sum_{j\ne i}a_{ij}x_j. Taking absolute values, |\lambda-a_{ii}||x_i| \le \sum_{j\ne i}|a_{ij}||x_j| \le \left(\sum_{j\ne i}|a_{ij}|\right)|x_i|. Since x_i\ne0, |\lambda-a_{ii}| \le \sum_{j\ne i}|a_{ij}|. Thus every eigenvalue lies in a Gershgorin disk.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.29

Problem formulation.

State and prove the refinement of Gershgorin’s theorem for connected components of the union of disks.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Let Ax=\lambda x, and choose i such that |x_i|=\max_j|x_j|. Then (\lambda-a_{ii})x_i=\sum_{j\ne i}a_{ij}x_j. Taking absolute values, |\lambda-a_{ii}||x_i| \le \sum_{j\ne i}|a_{ij}||x_j| \le \left(\sum_{j\ne i}|a_{ij}|\right)|x_i|. Since x_i\ne0, |\lambda-a_{ii}| \le \sum_{j\ne i}|a_{ij}|. Thus every eigenvalue lies in a Gershgorin disk.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.30

Problem formulation.

Prove the referenced result.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Assume A has eigenpairs (\lambda_j,v_j) with |\lambda_1|>|\lambda_2|\ge\cdots. Write x_0=\sum_j c_jv_j, \qquad c_1\ne0. Then A^kx_0=\lambda_1^k \left[ c_1v_1+\sum_{j\ge2}c_j \left(\frac{\lambda_j}{\lambda_1}\right)^k v_j \right]. After normalization, the terms with j\ge2 vanish geometrically at rate \left|\frac{\lambda_2}{\lambda_1}\right|^k. Thus the power method converges to the dominant eigenvector.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The order claim should be confirmed by a Taylor expansion or a log-log refinement table using a problem with a known exact solution.

Exercise 6.31

Problem formulation.

Give examples where the power method fails or stagnates.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Assume A has eigenpairs (\lambda_j,v_j) with |\lambda_1|>|\lambda_2|\ge\cdots. Write x_0=\sum_j c_jv_j, \qquad c_1\ne0. Then A^kx_0=\lambda_1^k \left[ c_1v_1+\sum_{j\ge2}c_j \left(\frac{\lambda_j}{\lambda_1}\right)^k v_j \right]. After normalization, the terms with j\ge2 vanish geometrically at rate \left|\frac{\lambda_2}{\lambda_1}\right|^k. Thus the power method converges to the dominant eigenvector.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.32

Problem formulation.

Prove the referenced result.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.33

Problem formulation.

Derive Rayleigh quotient iteration from shifted inverse iteration.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

For A=A^T, the Rayleigh quotient is \rho(x)=\frac{x^TAx}{x^Tx}. If x is an eigenvector, Ax=\lambda x, then \rho(x)=\frac{x^T\lambda x}{x^Tx}=\lambda. The stationary points of \rho(x) on the unit sphere satisfy Ax=\lambda x, so eigenvectors are precisely stationary points of the Rayleigh quotient.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.34

Problem formulation.

Prove the referenced result.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.35

Problem formulation.

Show that a shifted QR step also produces a unitarily similar matrix.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.36

Problem formulation.

Show that QR iteration preserves upper Hessenberg form.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.37

Problem formulation.

Derive the plane rotation that zeros the (p,q) entry of a real symmetric matrix.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.38

Problem formulation.

Derive the Arnoldi relation AV_k=V_kH_k+h_{k+1,k}v_{k+1}e_k^T.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.39

Problem formulation.

Prove the referenced result.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.40

Problem formulation.

Derive the Lanczos three-term recurrence from Arnoldi when A=A^*.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.41

Problem formulation.

Prove the referenced result in full detail.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.42

Problem formulation.

State and prove both min-max and max-min forms of the Courant–Fischer theorem.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.43

Problem formulation.

Prove the Cauchy interlacing theorem for eigenvalues of Hermitian matrices.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.44

Problem formulation.

State and prove Weyl’s eigenvalue perturbation inequality for Hermitian matrices.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.45

Problem formulation.

State the Davis–Kahan sin-theta theorem and explain its meaning for invariant subspace perturbation.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.46

Problem formulation.

Prove the referenced result.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.47

Problem formulation.

For Hermitian A, derive an angle bound between an approximate eigenvector and an exact eigenspace using a spectral gap.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.48

Problem formulation.

Prove the referenced result.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The order claim should be confirmed by a Taylor expansion or a log-log refinement table using a problem with a known exact solution.

Exercise 6.49

Problem formulation.

Prove the referenced result.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.50

Problem formulation.

Prove the equivalence between pseudospectra defined by resolvent norm and by eigenvalues of perturbed matrices.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.51

Problem formulation.

Construct a non-normal matrix whose eigenvalues are highly sensitive to tiny perturbations.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.52

Problem formulation.

Show that for Hermitian matrices, the Rayleigh quotient eigenvalue error is second order in the eigenvector angle.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

For A=A^T, the Rayleigh quotient is \rho(x)=\frac{x^TAx}{x^Tx}. If x is an eigenvector, Ax=\lambda x, then \rho(x)=\frac{x^T\lambda x}{x^Tx}=\lambda. The stationary points of \rho(x) on the unit sphere satisfy Ax=\lambda x, so eigenvectors are precisely stationary points of the Rayleigh quotient.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The order claim should be confirmed by a Taylor expansion or a log-log refinement table using a problem with a known exact solution.

Exercise 6.53

Problem formulation.

Prove the referenced result more rigorously.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The order claim should be confirmed by a Taylor expansion or a log-log refinement table using a problem with a known exact solution.

Exercise 6.54

Problem formulation.

Analyze convergence of orthogonal iteration to a dominant invariant subspace.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The order claim should be confirmed by a Taylor expansion or a log-log refinement table using a problem with a known exact solution.

Exercise 6.55

Problem formulation.

Show the relationship between QR iteration and orthogonal iteration.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.56

Problem formulation.

Derive the operation count for Householder reduction to Hessenberg form.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.57

Problem formulation.

Derive Householder tridiagonalization for real symmetric matrices.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.58

Problem formulation.

Derive the Wilkinson shift for a symmetric tridiagonal QR step.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.59

Problem formulation.

Explain deflation in the QR algorithm and derive a practical deflation criterion.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.60

Problem formulation.

Describe the implicit Francis double-shift QR step for real nonsymmetric matrices.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.61

Problem formulation.

Explain how loss of orthogonality in Lanczos can produce ghost eigenvalues.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.62

Problem formulation.

Describe selective reorthogonalization in Lanczos methods.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.63

Problem formulation.

Explain the need for restarting in Arnoldi eigenvalue computations.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.64

Problem formulation.

Study the implicitly restarted Arnoldi method and its relation to shifted QR.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.65

Problem formulation.

Analyze shift-and-invert Arnoldi for computing interior eigenvalues.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.66

Problem formulation.

Study the backward stability of the shifted QR algorithm and explain the role of unitary transformations.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

Always test the result on the scalar model equation or Fourier mode and verify the amplification factor satisfies the stated stability bound.

Exercise 6.67

Problem formulation.

Investigate aggressive early deflation in modern QR eigensolvers.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.68

Problem formulation.

Study divide-and-conquer methods for symmetric tridiagonal eigenvalue problems.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.69

Problem formulation.

Study the MRRR algorithm for symmetric tridiagonal eigenproblems and discuss its accuracy tradeoffs.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.70

Problem formulation.

Analyze why Jacobi-type methods can deliver high relative accuracy for certain symmetric eigenvalue problems.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.71

Problem formulation.

Design a Chebyshev polynomial filter for separating eigenvalues in a target interval.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.72

Problem formulation.

Study rational Krylov subspaces and compare them with polynomial Krylov subspaces.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.73

Problem formulation.

Derive the Jacobi–Davidson correction equation and explain how it targets selected eigenpairs.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.74

Problem formulation.

Study the locally optimal block preconditioned conjugate-gradient method for Hermitian eigenproblems.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

Distinguish forward error from backward error; a small residual does not imply a small forward error when the problem is ill-conditioned.

Exercise 6.75

Problem formulation.

Explain the FEAST algorithm as a contour-integration spectral projector method.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.76

Problem formulation.

Analyze the problem Ax=\lambda Bx, \qquad A=A^*, \qquad B=B^*>0. Derive a reduction to a standard Hermitian eigenvalue problem.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.77

Problem formulation.

Study the QZ algorithm for generalized eigenvalue problems and compare it with QR.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.78

Problem formulation.

Study polynomial eigenvalue problems and linearization strategies.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.79

Problem formulation.

Investigate nonlinear eigenvalue problems T(\lambda)x=0 and discuss Newton-type methods for them.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

Verify the stationarity residual, feasibility residual, and descent or curvature condition, not only the final numerical value.

Exercise 6.80

Problem formulation.

Design a numerical method for plotting pseudospectra using resolvent norms.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.81

Problem formulation.

Analyze eigenvalue computations for large sparse matrices arising from finite difference or finite element discretizations of elliptic operators.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.82

Problem formulation.

Study randomized subspace iteration for approximating dominant eigenspaces.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.83

Problem formulation.

Investigate communication-avoiding variants of Arnoldi or Lanczos methods.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.84

Problem formulation.

Analyze the use of mixed precision in eigenvalue computations, including residual correction and refinement.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Exercise 6.85

Problem formulation.

Develop interval or residual-based certified enclosures for Hermitian eigenvalues.

Method.

Use spectral decomposition, Rayleigh quotient identities, QR/inverse iteration, residual eigenvalue bounds, and perturbation theory.

Detailed solution and justification.

Use the eigenvalue relation Ax=\lambda x and analyze the method through the spectral decomposition of A. Residuals provide a posteriori information: r=Ax-\mu x. For Hermitian A, if \mu=\rho(x), then \operatorname{dist}(\mu,\sigma(A))\le \|r\|_2. For diagonalizable matrices, perturbation bounds involve the conditioning of the eigenvector matrix: A=V\Lambda V^{-1}, \qquad \kappa(V)=\|V\|\|V^{-1}\|.

Conclusion.

The displayed derivation gives the requested formula, proof, or computation. The final expression should be checked against the assumptions and notation of the corresponding chapter.

Diagnostic comment.

The formula should be verified by substitution into the defining equation and, when possible, by a small numerical test.

Interactive tools

The theory and solutions come first. These tools are for review, laboratories, games, memory cards, and randomized assessment.

Professional visualsAnimated diagrams and conceptual graphics. Programming practiceDebug JavaScript/PHP method snippets with corrected code.

Interactive PHP laboratories for this chapter

These experiments run inside the same web page. The student changes the input data and presses Run PHP. Results are returned by the PHP server and displayed immediately.