× Binary Unit Round-off Roundoff Error Number System with base \(\beta\) Floating point error analysis Orders of Convergence Newton's Method Error Analysis for Newton's Method Theorems on Newton's Method Horner's Algorithm Contractive Mapping Polynomial Interpretation Hermite Interpolation Change of Intervals Gaussian Quadrature Error Analysis for Gaussian Quadrature Composite Formulae \(\theta\)-Method Error Initial Value Problem Stability Local Truncation Error for Multistep Methods Schur's Form Hermite Matrix QR Factorization Householder LU Decomposition Norm Positive Definite Gerschgorin Theorem Rayleigh Quotient August 2017 August 2018 January 2019 August 2019 January 2020 August 2020 January 2021 August 2021 January 2022 August 2022 January 2023 August 2023 January 2024 References
☰ Menu

My name is Hanzhang Yin, and I have developed this website as a resource to facilitate the review of key concepts in abstract algebra, and it is not fully completed. Feel free to email me at hanyin@ku.edu if there are any errors or suggestions for improvement.

2023 (August) Numerical Analysis Qualifying Examination

1. Suppose that \( p \) is a double zero of the function \( f \). Thus \( f(p) = f'(p) = 0 \ne f''(p) \). Show that if \( f'' \) is continuous, then in Newton's method we shall have \[ \lim_{n \to \infty} \frac{|p - p_n|}{|p - p_{n-1}|} = \frac{1}{2}. \]

Proof. Suppose that \(p - x_{n-1} = e\), then we can know that \[ \begin{align} f(x_{n-1}) &= f(p - e) = f(p) + f'(p)(-e) + \frac{f''(p)(-e)^2}{2} + \mathcal{O}(e^3) = \frac{f''(p)e^2}{2} + \mathcal{O}(e^3), \\ f'(x_{n-1}) &= f'(p) + f''(p)(-e) + \mathcal{O}(e^2) = -f''(p)e + \mathcal{O}(e^2). \end{align} \] According to the Newton's method, we have \(x_{n} = x_{n-1} - \dfrac{f(x_{n-1})}{f'(x_{n-1})}\). Then, we can know that \[ \begin{align} p - x_{n} &= p - x_{n-1} + \frac{f(x_{n-1})}{f'(x_{n-1})} \\ &= e + \dfrac{f(x_{n-1})}{f'(x_{n-1})} = e + \dfrac{\frac{f''(p)e^2}{2} + \mathcal{O}(e^3)}{-f''(p)e + \mathcal{O}(e^2)} = \dfrac{-f''(p)e^2 + \mathcal{O}(e^3)}{-f''(p)e + \mathcal{O}(e^2)} + \dfrac{\frac{f''(p)e^2}{2} + \mathcal{O}(e^3)}{-f''(p)e + \mathcal{O}(e^2)}, \\ &= \dfrac{-\frac{f''(p)e^2}{2} + \mathcal{O}(e^3)}{f''(p)e + \mathcal{O}(e^2)}, \\ \dfrac{p - x_{n}}{p - x_{n-1}} &= \dfrac{-\frac{f''(p)e^2}{2} + \mathcal{O}(e^3)}{-f''(p)e^2 + \mathcal{O}(e^3)}, \\ &= \dfrac{(-\frac{f''(p)e^2}{2} + \mathcal{O}(e^3))/e^2}{(-f''(p)e^2 + \mathcal{O}(e^3))/e^2} \\ &= \dfrac{-\frac{f''(p)}{2} + \mathcal{O}(e)}{-f''(p) + \mathcal{O}(e)}.\\ \left|\dfrac{p - x_{n}}{p - x_{n-1}}\right| &= \left|\dfrac{-\frac{f''(p)}{2} + \mathcal{O}(e)}{-f''(p) + \mathcal{O}(e)}\right|. \end{align} \] Since \(e\to 0\) when \(n\to \infty\), we have \[ \begin{align} \lim_{n\to\infty}\left|\dfrac{p - x_{n}}{p - x_{n-1}}\right| &= \lim_{n\to\infty}\left|-\dfrac{\frac{f''(p)}{2} + \mathcal{O}(e)}{-f''(p) + \mathcal{O}(e)}\right|, \\ &= \dfrac{\frac{f''(p)}{2}}{f''(p)}\\ &= \frac{1}{2}. \tag*{\(\blacksquare\)} \end{align} \]

2. (a) Consider a numerical integration rule of the form \[ \int_{-1}^1 f(x) \, dx \approx A f \left( -\sqrt{\frac{3}{5}} \right) + B f(0) + C f \left( \sqrt{\frac{3}{5}} \right) \] What is the linear system that must be solved in order to determine the coefficients \( A, B, C \) so that the formula has a possibly highest degree of precision? Solve for \( A, B, C \). What is the degree of precision?

(b) Find the error term for the rule.

(c) Derive the composite quadrature rule resulting from the application of the rule in Question (2a) to the general integral \( \int_a^b f(x) \, dx \) on a uniform partition \( x_0 = a \lt x_1 \lt \cdots \lt x_n = b \).

(d) Derive the error term for the composite rule. It is required that the error term does not contain summations.

Solution (a). To determine the coefficients \( A, B, C \), we need to solve the following linear system: \[ \begin{align} \int_{-1}^1 1 \, dx &= A + B + C, \\ \int_{-1}^1 x \, dx &= -\sqrt{\frac{3}{5}} A + 0 + \sqrt{\frac{3}{5}} C, \\ \int_{-1}^1 x^2 \, dx &= \left( -\sqrt{\frac{3}{5}} \right)^2 A + 0^2 + \left( \sqrt{\frac{3}{5}} \right)^2 C. \end{align} \] Solving the system, we have \[ \begin{align} A + B + C &= 2, \\ -\sqrt{\frac{3}{5}} A + \sqrt{\frac{3}{5}} C &= 0, \\ \frac{3}{5} A + \frac{3}{5} C &= \frac{2}{3}. \end{align} \] From the second equation, we have \( A = C \). Substituting this into the third equation, we have \( A = C = \frac{5}{9} \). Then, we have \( B = \frac{8}{9} \). Therefore, we have the formula \[ \int_{-1}^1 f(x) \, dx \approx \frac{5}{9} f \left( -\sqrt{\frac{3}{5}} \right) + \frac{8}{9} f(0) + \frac{5}{9} f \left( \sqrt{\frac{3}{5}} \right). \] Then we try to find the degree of precision. \[ \begin{align} \int_{-1}^1 x^3 \, dx &= \frac{5}{9} \left( -\sqrt{\frac{3}{5}} \right)^3 + \frac{8}{9}\cdot 0 + \frac{5}{9} \left( \sqrt{\frac{3}{5}} \right)^3 = 0, \\ \int_{-1}^1 x^4 \, dx &= \frac{5}{9} \left( -\sqrt{\frac{3}{5}} \right)^4 + \frac{8}{9}\cdot 0 + \frac{5}{9} \left( \sqrt{\frac{3}{5}} \right)^4 = \frac{2}{5}, \\ \int_{-1}^1 x^5 \, dx &= \frac{5}{9} \left( -\sqrt{\frac{3}{5}} \right)^5 + \frac{8}{9}\cdot 0 + \frac{5}{9} \left( \sqrt{\frac{3}{5}} \right)^5 = 0, \\ \int_{-1}^1 x^6 \, dx = \frac{2}{7} &\neq \frac{5}{9} \left( -\sqrt{\frac{3}{5}} \right)^6 + \frac{8}{9}\cdot 0 + \frac{5}{9} \left( \sqrt{\frac{3}{5}} \right)^6 = \frac{5}{25}. \end{align} \] Thus, the degree of precision is \(5\).

Solution (b). We can know the error term is in the form of \(Cf^{(6)}(\xi)\) for some \(\xi\in(-1, 1)\) since the degree of precision is 5. Then, we have And we already found that \[ \int_{-1}^1 x^6 \, dx = \frac{2}{7} \neq \frac{5}{9} \left( -\sqrt{\frac{3}{5}} \right)^6 + \frac{8}{9}\cdot 0 + \frac{5}{9} \left( \sqrt{\frac{3}{5}} \right)^6 = \frac{5}{25}. \] Hence, we have \[ \begin{align} \int_{-1}^1 x^6 \, dx - \frac{5}{9} \left( -\sqrt{\frac{3}{5}} \right)^6 + \frac{8}{9}\cdot 0 + \frac{5}{9} \left( \sqrt{\frac{3}{5}} \right)^6 &= \frac{2}{7} - \frac{6}{25} = \frac{8}{175} = C f^{(6)}(\xi), \\ f^{(6)}(x) &= 720 \\ 720C &= \frac{8}{175} \\ C &= \frac{8}{175\cdot 720} = \frac{1}{15750}. \end{align} \] Thus, the error term is \( \frac{1}{15750}f^{(6)}(\xi) \) for some \( \xi \in (-1, 1) \).

Solution (c). We will apply the change of the interval from \([-1, 1]\) to \([x_i, x_{i+1}]\). Firstly, we define \(\lambda(t) = \dfrac{x_{i+1} - x_i}{2}t + \dfrac{x_{i+1} + x_i}{2}\) for \(t\in[-1, 1]\). Then, we have \(\lambda(1) = x_{i+1}\) and \(\lambda(-1) = x_i\). Hence, we have \[ \begin{align} \int_{x_i}^{x_{i+1}} f(x) \, dx &= \int_{-1}^1 f(\lambda(t))\lambda'(t) \, dt \\ &= \dfrac{x_{i+1} - x_i}{2} \int_{-1}^1 f\left(\dfrac{x_{i+1} - x_i}{2}t + \dfrac{x_{i+1} + x_i}{2}\right)\, dt, \\ &\approx \dfrac{x_{i+1} - x_i}{2} \left( \dfrac{5}{9} f\left( \dfrac{x_{i+1} - x_i}{2}\left(-\sqrt{\frac{3}{5}}\right) + \dfrac{x_{i+1} + x_i}{2} \right) + \dfrac{8}{9} f\left(\dfrac{x_{i+1} + x_i}{2}\right) + \dfrac{5}{9} f\left( \dfrac{x_{i+1} - x_i}{2}\sqrt{\frac{3}{5}} + \dfrac{x_{i+1} + x_i}{2}\right) \right). \end{align} \] Then, we have \[ \int_a^b f(x)\, dx \approx \sum_{i=0}^{n-1} \left[\dfrac{x_{i+1} - x_i}{2} \left( \dfrac{5}{9} f\left( \dfrac{x_{i+1} - x_i}{2}\left(-\sqrt{\frac{3}{5}}\right) + \dfrac{x_{i+1} + x_i}{2} \right) + \dfrac{8}{9} f\left(\dfrac{x_{i+1} + x_i}{2}\right) + \dfrac{5}{9} f\left( \dfrac{x_{i+1} - x_i}{2}\sqrt{\frac{3}{5}} + \dfrac{x_{i+1} + x_i}{2}\right) \right)\right]. \]

Solution (d). We already found that the error term is in the form of \( \frac{1}{15750}f^{(6)}(\xi) \) for some \( \xi \in (-1, 1) \) when we try to approximate \( \int_{-1}^1 f(x) \, dx \). Hence, we have error term for \[ \int_{x_i}^{x_{i+1}} f(x)\, dx = \dfrac{x_{i+1} - x_i}{2} \left( \dfrac{5}{9} f\left( \dfrac{x_{i+1} - x_i}{2}\left(-\sqrt{\frac{3}{5}}\right) + \dfrac{x_{i+1} + x_i}{2} \right) + \dfrac{8}{9} f\left(\dfrac{x_{i+1} + x_i}{2}\right) + \dfrac{5}{9} f\left( \dfrac{x_{i+1} - x_i}{2}\sqrt{\frac{3}{5}} + \dfrac{x_{i+1} + x_i}{2}\right) \right) + \dfrac{x_{i+1} - x_i}{31500}f^{(6)}\left(\dfrac{x_{i+1} - x_i}{2}\xi+ \dfrac{x_{i+1} - x_i}{2}\right). \] Hence, the error term for approximating \(\displaystyle \int_{x_i}^{x_{i+1}} f(x)\, dx\) is \[ E_i = \dfrac{x_{i+1} - x_i}{31500}f^{(6)}\left(\dfrac{x_{i+1} - x_i}{2}\xi+ \dfrac{x_{i+1} - x_i}{2}\right),\text{ for some } \xi\in(-1, 1). \] Given we have uniformly partitions, we can say that the error term for approximating \(\displaystyle \int_{x_i}^{x_{i+1}} f(x)\, dx\) is \[ E_i = \dfrac{h}{31500}f^{(6)}\left(\eta_i\right),\text{ for some } \eta_i\in(x_{i}, x_{i+1}). \] Suppose that \(f\) is smooth. In that case, the error term for the composite rule is \[ E = \sum_{i=0}^{n-1} E_i = \sum_{i=0}^{n-1} \dfrac{h}{31500}f^{(6)}\left(\eta_i\right) = \dfrac{h}{31500}\sum_{i=0}^{n-1} f^{(6)}\left(\eta_i\right) = \dfrac{h}{31500}f^{(6)}\left(\eta\right),\text{ for some } \eta\in(a, b). \]

\(\textbf{Problem 3. }\)Consider the \(\theta\)-method \[ y_{n+1} = y_n + h \left( \theta f_{n+1} + (1 - \theta) f_n \right), \quad \theta \in [0, 1], \] where \( f_n = f(t_n, y_n) \).

  • (a). For which value(s) of \(\theta\) is the \(\theta\)-method explicit? Implicit?
  • (b). Find an expression for the local truncation error in terms of \(\theta\), \(h\), and derivatives of the exact solution \(y(t)\) to the initial value problem. What are the local truncation errors for \(\theta = 0\), \(\frac{1}{2}\), \(1\)?

Solution (a). To make a method explicit, we only need values of \(y_n\) and \(f_n\). Hence, when \(\theta = 0\), we have \[ y_{n+1} = y_n + h f_n, \] which is explicit. To make a method implicit, we need values of \(f_{n+1}\). Hence, when \(\theta = 1\), we have \[ y_{n+1} = y_n + h f_{n+1}, \] which is implicit.

Solution (b). The local truncation error is given by \[\tau_{n+1} = y(t_{n+1}) - y_{n+1} - h(\theta f(t_{n+1}, y(t_{n+1})) + (1 - \theta) f(t_n, y(t_n))).\] When \(\theta = 0\), we have \[\tau_{n+1} = y(t_{n+1}) - y_n - h f(t_n, y(t_n)),\] which is the local truncation error for \(\theta = 0\). When \(\theta = \frac{1}{2}\), we have \[\tau_{n+1} = y(t_{n+1}) - y_n - \frac{h}{2} (f(t_{n+1}, y(t_{n+1})) + f(t_n, y(t_n))),\] which is the local truncation error for \(\theta = \frac{1}{2}\). When \(\theta = 1\), we have \[\tau_{n+1} = y(t_{n+1}) - y_n - h f(t_{n+1}, y(t_{n+1})),\] which is the local truncation error for \(\theta = 1\).

4. Suppose \(A\in\mathbb{C}^{n\times n}\). Let \(\sigma_1\) and \(\sigma_n\) be the largest and smallest singular values of \(A\), respectively.

(a) Prove \(\|Ax\|_2\geq \sigma_n\|x\|_2\) for any \(x\in \mathbb{C}^n\).

(b) Prove that if \(B\in \mathbb{C}^{n\times n}\) is singular, then \(\|A - B\|_2\geq \sigma_n\).

(c) Show there exists a singular matrix \(B\in \mathbb{C}^{n\times n}\) such that \(\|A - B\|_2 = \sigma_n\).

(d) Let \(\lambda\) be an eigenvalue of \(A\). Prove \[ \sigma_n\leq |\lambda_n|\leq \sigma_1. \]

Proof (a). Firstly, we denote \(A\) as SVD such as \(A = U\Sigma V^*\). Hence, we will have \[ \|Ax\|_2 = \|U\Sigma V^* x\|_2 = \|\Sigma V^* x\|_2, \] for \(U\) is unitary. We denote \(V^*x = v = \begin{bmatrix} v_1 \\ \vdots \\ v_n\end{bmatrix}\). Then, we can get \[ \|Ax\|_2 = \|\Sigma v\|_2 = \left\|\begin{bmatrix} \sigma_1\cdot v_1 \\ \vdots \sigma_n\cdot v_n \end{bmatrix}\right\|_2 = \sqrt{(\sigma_1^2\cdot v_1\cdot \overline{v_1} + \dots + \sigma_n^2\cdot v_n\cdot \overline{v_n}} = = \sqrt{(\sigma_1^2\cdot \|v_1\|_2^2 + \dots + \sigma_n^2\cdot \|v_n\|^2} \geq \sqrt{\sigma_n^2\cdot \|v_1\|^2 + \dots + \sigma_n^2\cdot \|v_n\|^2} = \sigma_n\|v\|_2. \] Since \(V\) is unitary, we can know that \(V^*\) is unitary and \(\|v\|_2 = \|V^* x\|_2 = \|x\|_2\). Therefore, we can conclude that \[ \|Ax\|_2\geq \sigma_n\|v\|_2 = \sigma_n\|x\|_2. \blacksquare \]

Proof (b). Given that \(B\) is singular, it contains an eigenvalue of \(0\). Let \(v\) be the eigenvector correspondent to \(0\) where \(\|v\|_2 = 1\). Hence, we have \[ \|A - B\|_2 = \max_{\|x\|=1}\|(A - B)x\|_2 \geq \|(A - B)v\|_2 = \|Av - Bv\|_2 = \|Av - 0\cdot v\|_2 = \|Av\|_2. \] We will borrow the notation from the previous part: \(A = U\Sigma V^*\). Then, we have \[ \|Av\|_2 = \|U\Sigma V^* v\|_2 = \|\Sigma V^* v\|_2. \] We denote \(u = V^*v = \begin{bmatrix} u_1 \\ \vdots \\ u_n \end{bmatrix}\). Thus, we have \(\|u\|_2 = \|V^*v\|_2 = \|v\|_2 = 1\) for \(V\) is unitary. \[ \|\Sigma V^*v\|_2 = \|\Sigma u\|_2 = \sqrt{(\sigma_1^2\cdot u_1\cdot \overline{u_1} + \dots + \sigma_n^2 u_n\cdot \overline{u_n})} = \sqrt{(\sigma_1^2\cdot \|u_1\|_2^2 + \dots + \sigma_n^2 \|u_n\|_2^2)} \geq \sqrt{(\sigma_n^2\cdot \|u_1\|_2^2 + \dots + \sigma_n^2 \|u_n\|_2^2)} = \sigma_n\|u\|_2 = \sigma_n. \] Thus, we can know that \(\|A-B\|_2\geq \sigma_n\). \(\blacksquare\)

Proof (c). Given that \(A = U\Sigma V^*\) where \[ \Sigma = \begin{bmatrix} \sigma_1 & 0 & \dots & 0 \\ 0 & \sigma_2 & \dots & 0 \\ \vdots & 0 & \ddots & 0 \\ 0 & \dots & 0 & \sigma_n \\ \end{bmatrix}. \] Let \(B = U\Sigma' V^*\) where \[ \Sigma' = \begin{bmatrix} \sigma_1 & 0 & \dots & 0 \\ 0 & \sigma_2 & \dots & 0 \\ \vdots & 0 & \ddots & 0 \\ 0 & \dots & 0 & 0 \\ \end{bmatrix}. \] In that case, we can know that \(\det(B) = \det(U)\det(\Sigma')\det(V^*) = 1\cdot 0\cdot 1 = 0\), which implies that \(B\) is singular. Then, we can get \[ \|A - B\|_2 = \|U\Sigma V^* - U\Sigma' V^*\|_2 = \|U(\Sigma - \Sigma')V^*\|_2 = \|\Sigma - \Sigma'\|_2 = \left\|\begin{bmatrix} 0 & 0 & \dots & 0 \\ 0 & 0 & \dots & 0 \\ \vdots & 0 & \ddots & 0 \\ 0 & \dots & 0 & \sigma_n \\ \end{bmatrix}\right\|_2 \] Let \(v = \begin{bmatrix} v_1 \\ \vdots \\ v_n \end{bmatrix} \in \mathbb{C}^n\) such that \(\|v\|_2 = 1\) and \(\|Av\|_2 = \max\limits_{\|x\|_2=1}\|Ax\|_2\). Then, we can get \[ \begin{align} \|A - B\|_2 &= \left\|\begin{bmatrix} 0 & 0 & \dots & 0 \\ 0 & 0 & \dots & 0 \\ \vdots & 0 & \ddots & 0 \\ 0 & \dots & 0 & \sigma_n \\ \end{bmatrix}\right\|_2 = \max_{\|x\|_2=1}\left\|\begin{bmatrix} 0 & 0 & \dots & 0 \\ 0 & 0 & \dots & 0 \\ \vdots & 0 & \ddots & 0 \\ 0 & \dots & 0 & \sigma_n \\ \end{bmatrix}\cdot x\right\|_2 = \left\|\begin{bmatrix} 0 & 0 & \dots & 0 \\ 0 & 0 & \dots & 0 \\ \vdots & 0 & \ddots & 0 \\ 0 & \dots & 0 & \sigma_n \end{bmatrix}\cdot \begin{bmatrix} v_1 \\ \vdots \\ v_n \end{bmatrix}\right\|_2 \\ &= \left\|\begin{bmatrix} 0 \\ \vdots \\ \sigma_nv_n \end{bmatrix}\right\|_2 = \sigma_n\|v_n\|_2 \leq \sigma_n(\sqrt{\|v_1\|_2^2 + \dots + \|v_n\|_2^2}) = \sigma_n. \end{align} \] Hence, we have \(\|A- B\|_2\leq \sigma_n\). Now, let \(u = \vec{e_n}\). \[ \|A - B\|_2 = \left\|\begin{bmatrix} 0 & 0 & \dots & 0 \\ 0 & 0 & \dots & 0 \\ \vdots & 0 & \ddots & 0 \\ 0 & \dots & 0 & \sigma_n \\ \end{bmatrix}\right\|_2 = \max_{\|x\|=1}\left\|\begin{bmatrix} 0 & 0 & \dots & 0 \\ 0 & 0 & \dots & 0 \\ \vdots & 0 & \ddots & 0 \\ 0 & \dots & 0 & \sigma_n \\ \end{bmatrix}\cdot x\right\|_2 \geq \left\|\begin{bmatrix} 0 & 0 & \dots & 0 \\ 0 & 0 & \dots & 0 \\ \vdots & 0 & \ddots & 0 \\ 0 & \dots & 0 & \sigma_n \\ \end{bmatrix}\cdot \begin{bmatrix} 0 \\ \vdots \\ 1 \end{bmatrix}\right\|_2 = \left\|\begin{bmatrix} 0 \\ \vdots \\ \sigma_n \end{bmatrix}\right\|_2 = \sigma_n. \] Therefore, we have \(\|A - B\|_2 = \sigma_n\). \(\blacksquare\)

Proof (d). Since \(\lambda\) is the eigenvalue of \(A\), we denote \(v\) as the eigenvector of \(A\) which is correspondent to \(\lambda\) and \(\|v\|_2 = 1\). From part (a), we can know that \(\|Ax\|_2\geq \sigma_n\|x\|_2\). Thus, we have \[ |\lambda| = |\lambda|\|v\|_2 = \|\lambda v\|_2 = \|Av\|_2 \geq \sigma_n\|v\|_2 = \sigma_n. \] For the other direction, we have \[ \sigma_1 = \|A\|_2 = \max_{\|x\|=1} \|Ax\|_2 \geq \|Av\|_2 = |\lambda|\|v\|_2 = |\lambda|. \] Therefore, we have \[ \sigma_n\leq |\lambda|\leq \sigma_1. \blacksquare \]

6. Let \( A \in \mathbb{C}^{m \times m} \), and \( A = U\Sigma V^* \) being a singular value decomposition of \( A \), where \( U, V \in \mathbb{C}^{m \times m} \) are unitary and \( \Sigma \) is non-negative diagonal. Define \[ W = \frac{1}{\sqrt{2}} \begin{bmatrix} U & U \\ V & -V \end{bmatrix} \in \mathbb{C}^{2m \times 2m}. \]

(a). Prove \( W \) is unitary.

(b). Let \[ B = \begin{bmatrix} 0 & A \\ A^* & 0 \end{bmatrix} \in \mathbb{C}^{2m \times 2m}. \] Prove \( B \) is Hermitian and have the spectral decomposition \[ B = W \begin{bmatrix} \Sigma & 0 \\ 0 & -\Sigma \end{bmatrix} W^*. \] Prove \( \|B\|_2 = \|A\|_2 \).

(c). Let \( \tilde{\sigma} \) be a singular value of \( \tilde{A} = A + E \), where \( E \in \mathbb{C}^{m \times m} \). Prove there exists a singular value \( \sigma \) of \( A \) such that \[ |\tilde{\sigma} - \sigma| \leq \|E\|_2. \]

Proof (a). Given \(W\), we have with Unitary Matrices \(U, V\), we can get \[ W^* = \frac{1}{\sqrt{2}} \begin{bmatrix} U^* & V^* \\ U^* & -V^* \end{bmatrix} \in \mathbb{C}^{2m \times 2m}. \] Then we have \begin{align} W^* W &= \frac{1}{2} \begin{bmatrix} U^*U + V^*V & U^*U - V^*V \\ U^*U - V^*V & U^*U + V^*V \end{bmatrix}\\ &= \frac{1}{2} \begin{bmatrix} I_m + I_m & I_m - I_m \\ I_m - I_m & I_m + I_m \end{bmatrix}\\ &= \frac{1}{2} \begin{bmatrix} 2I_m & 0_m \\ 0_m & 2I_m \end{bmatrix}\\ &= \begin{bmatrix} I_m & 0_m \\ 0_m & I_m \end{bmatrix} \\ &= I_{2m}. \end{align} By the similar process, we have \begin{align} WW^* &= \frac{1}{2} \begin{bmatrix} UU^* + UU^* & UV^* - UV^* \\ VU^* - VU^* & VV^* + VV^* \end{bmatrix}\\ &= \frac{1}{2} \begin{bmatrix} I_m + I_m & 0_m \\ 0_m & I_m + I_m \end{bmatrix}\\ &= I_{2m}. \end{align} Therefore, we show that \(W\) is unitary. \[ \tag*{$\square$} \]

Proof (b).Firstly, we have \[ B^* = \begin{bmatrix} 0 & A^* \\ (A^*)^* & 0 \end{bmatrix} = \begin{bmatrix} 0 & A^* \\ A & 0 \end{bmatrix} = B. \] Thus, we have \(B\) is Hermitian. To show the spectral decomposition, we need to do calculation. \begin{align} W \begin{bmatrix} \Sigma & 0 \\ 0 & -\Sigma \end{bmatrix} W^* &= \frac{1}{2} \begin{bmatrix} U & U \\ V & -V \end{bmatrix} \begin{bmatrix} \Sigma & 0 \\ 0 & -\Sigma \end{bmatrix} \begin{bmatrix} U^* & V^* \\ U^* & -V^* \end{bmatrix}\\ &= \frac{1}{2} \begin{bmatrix} U\Sigma & -U\Sigma \\ V\Sigma & V\Sigma \end{bmatrix} \begin{bmatrix} U^* & V^* \\ U^* & -V^* \end{bmatrix}\\ &= \frac{1}{2} \begin{bmatrix} U\Sigma U^* - U\Sigma U^* & U\Sigma V^* + U\Sigma V^* \\ V\Sigma U^* + V\Sigma U^* & V\Sigma V^* - V\Sigma V^* \end{bmatrix}\\ &= \begin{bmatrix} 0 & U\Sigma V^* \\ V\Sigma U^* & 0 \end{bmatrix}\\ \end{align} Since \(A = U\Sigma V^*\), which is SVD of \(A\), we know all entries of \(\Sigma\) are real. Thus, we have \(\Sigma = \Sigma^*\). In that case, we have \[ W \begin{bmatrix} \Sigma & 0 \\ 0 & -\Sigma \end{bmatrix} W^* = \begin{bmatrix} 0 & U\Sigma V^* \\ V\Sigma U^* & 0 \end{bmatrix} = \begin{bmatrix} 0 & U\Sigma V^* \\ V\Sigma^* U^* & 0 \end{bmatrix} = \begin{bmatrix} 0 & A \\ A^* & 0 \end{bmatrix} = B. \] \[ \tag*{$\square$} \]

Since we just show that \(W\) is unitary, and we are given that \(U\) and \(V\) are unitary, we have \begin{align} \|A\|_2 &= \|U\Sigma V^*\|_2 = \|\Sigma\|_2 = \max\sigma_i \\ \|B\|_2 &= \left\|W \begin{bmatrix} \Sigma & 0 \\ 0 & -\Sigma \end{bmatrix} W^*\right\|_2 = \left\|\begin{bmatrix} \Sigma & 0 \\ 0 & -\Sigma \end{bmatrix}\right\|_2. \end{align} Since \(\begin{bmatrix} I_m & 0 \\ 0 & -I_m \end{bmatrix}\) is a unitary matrix, and \[ \begin{bmatrix} I_m & 0 \\ 0 & -I_m \end{bmatrix} \begin{bmatrix} \Sigma & 0 \\ 0 & -\Sigma \end{bmatrix} = \begin{bmatrix} \Sigma_m & 0 \\ 0 & \Sigma_m \end{bmatrix}, \] we have \[ \left\|\begin{bmatrix} \Sigma & 0 \\ 0 & -\Sigma \end{bmatrix}\right\|_2 = \left\|\begin{bmatrix} I_m & 0 \\ 0 & -I_m \end{bmatrix} \begin{bmatrix} \Sigma & 0 \\ 0 & -\Sigma \end{bmatrix}\right\|_2 = \left\|\begin{bmatrix} \Sigma_m & 0 \\ 0 & \Sigma_m \end{bmatrix}\right\|_2 = \max\sigma_i. \] Therefore, we have \(\|B\|_2 = \|A\|_2\). \[ \tag*{$\square$} \]

Proof (c).We already knew that \[ B = W \begin{bmatrix} \Sigma & 0 \\ 0 & -\Sigma \end{bmatrix} W^*. \] Let us define \(\tilde{B}\) such as \[ \begin{align} \tilde{B} &:= \begin{bmatrix} 0 & \tilde{A} \\ (\tilde{A})^* & 0 \end{bmatrix} \\ &= \begin{bmatrix} 0 & A + E \\ A^* + E^* & 0 \end{bmatrix} \\ &= \begin{bmatrix} 0 & A \\ A^* & 0 \end{bmatrix} + \begin{bmatrix} 0 & E \\ E^* & 0 \end{bmatrix} \\ &= W \begin{bmatrix} \Sigma & 0 \\ 0 & -\Sigma \end{bmatrix} W^* + \begin{bmatrix} 0 & E \\ E^* & 0 \end{bmatrix} \\ \end{align} \] We denote \(\begin{bmatrix} 0 & E \\ E^* & 0 \end{bmatrix}\) as \(F\). According to part (a), we can know that \(\tilde{A}\) has SVD \(\tilde{U}\tilde{\Sigma}\tilde{V}^*\) and \[ \tilde{A} = \tilde{W}\begin{bmatrix} \tilde{\Sigma} & 0 \\ 0 & -\tilde{\Sigma} \end{bmatrix}\tilde{W}^*, \] where \(\tilde{W}\) is unitary. According to \(\textbf{Bauer-Fiker Theorem}\), we can know that if \(\tilde{\sigma}\) is an eigenvalue of \(\tilde{B}\), there is an eigenvalue \(\sigma\) of \(B\) such that \[ |\tilde{\mu} - \mu|\leq \kappa(\tilde{W})\|\|F\|_2. \] According to part (b), we can know that \(\|F\|_2 = \|E\|_2\). Hence, we have \[ |\tilde{\mu} - \mu|\leq \kappa(\tilde{W})\|\|E\|_2 = \|\tilde{W}\|_2\|\tilde{W}^{-1}\|_2\|E\|_2 = \|E\|_2, \] for \(\tilde{W}\) is unitary. We know that \(\tilde{\mu}\) is \in \(\Lambda(\tilde{B}\), so it is an entry of \(\tilde{\Sigma}\) or \(-\tilde{\Sigma}\). By \(\textbf{ If \(\mu\geq 0\), then we can know that \(\tilde{\mu} = \tilde{\sigma}\). Then, we can know that \[ |\tilde{\sigma} - \mu|\leq \|E\|_2. \] After that, we will discuss the case that \(\mu\geq 0\). Similarly, we can know that \(\mu = \sigma\). Hence, we have \[ |\tilde{\sigma} - \sigma|\leq \|E\|_2. \] If \(\mu\lt 0\), then we can know that \(\mu = -\sigma\) for some singular value \(\sigma\). Hence, we have \[ |\tilde{\sigma} - \sigma|\leq|\tilde{\sigma} + \sigma|\leq \|E\|_2. \] for both \(\tilde{\sigma}\) and \(\sigma\) are positive. If \(\tilde{\mu}\lt 0\), then we can know that \(\tilde{\mu} = -\tilde{\sigma}\). Hence, we have \[ |\tilde{\mu} - \mu| = | -\tilde{\sigma} - \mu| = |\tilde{\sigma} + \mu|\leq \|E\|_2. \] If \(\mu\geq 0\), then we can know that \(\mu = \sigma\) for some \(\sigma\). Hence, we have \[ |\tilde{\sigma} - \sigma|< |\tilde{\sigma} + \sigma| =|\tilde{\sigma} + \mu|\leq \|E\|_2, \] for both \(\tilde{\sigma}\) and \(\sigma\) are positive. If \(\mu\lt 0\), then we can know that \(\mu = -\sigma\) for some singular value \(\sigma\). Hence, we have \[ |\tilde{\sigma} - \sigma| = |\tilde{\sigma} + \mu|\leq \|E\|_2. \tag*{\(\blacksquare\)} \]