41.8 Dilations

Definition 41.8.1 (Compression).label Let $H$ be a complex Hilbert space, $K \subset H$ be a closed subspace, $P \in B(H)$ be the orthogonal projection onto $K$, $S \in B(K)$, and $T \in B(H)$, then $S$ is a compression of $T$ if $S = PT|_{K}$.

Definition 41.8.2 (Dilation).label Let $H$ be a complex Hilbert space and $T \in B(H)$, then a dilation of $T$ is a pair $(K, S)$ where:

  1. (1)

    $K$ is a Hilbert space containing $H$.

  2. (2)

    $S \in B(K)$.

  3. (3)

    If $P \in B(K)$ is the orthogonal projection onto $H$, then $T = PS|_{H}$.

If $T^{n} = PS^{n}|_{H}$ for all $n \in \natz$, then $S$ is a power dilation of $T$.

Lemma 41.8.3.label Let $H$ be a complex Hilbert space, $T \in B(H)$, $(K, S)$ be a power dilation of $T$, and $P \in B(K)$ be the orthogonal projection onto $H$, then

  1. (1)

    $(K, S^{*})$ is a power dilation of $T^{*}$.

  2. (2)

    For any $p, q \in \complex[z]$,

    \[p(T) + q(T)^{*} = P[p(S) + q(S)^{*}]|_{H}\]

Proof. (1): For any $\xi, \eta \in H$ and $n \in \natp$, $P\xi = \xi$ and $P\eta = \eta$, so

\begin{align*}\dpn{(T^*)^n\xi, \eta}{H}&= \dpn{\xi, T^n\eta}{H}= \dpn{\xi, PS^n\eta}{K}= \dpn{P\xi, S^n\eta}{K}\\&= \dpn{(S^*)^nP\xi, P\eta}{K}= \dpn{P(S^*)^n\xi, \eta}{H}\end{align*}

$\square$

Lemma 41.8.4.label Let $H$ be a complex Hilbert space, $R \in B(H)$, $(K, S)$ be a dilation of $R$, and $(L, T)$ be a dilation of $S$, then $(L, T)$ is a dilation of $R$.

Similarly, if $(K, S)$ is a power dilation of $R$ and $(L, T)$ is a power dilation of $S$, then $(L, T)$ is a power dilation of $R$.

Theorem 41.8.5 (Stinespring’s Dilation Theorem).label Let $A$ be a unital $C^{*}$-algebra, $H$ be a complex Hilbert space, and $\phi: A \to B(H)$ be a completely positive map, then there exists a triple $(\pi, V, K)$, where:

  1. (a)

    $K$ is a complex Hilbert space.

  2. (b)

    $V: H \to K$ is a continuous linear map.

  3. (c)

    $\pi: A \to B(K)$ is a unital *-homomorphism.

such that:

  1. (1)

    $\norm{\phi(1_A)}_{B(H)}= \norm{V}_{L(H; K)}^{2}$.

  2. (2)

    For all $a \in A$, $\phi(a) = V^{*}\pi(a)V$.

Moreover, if $\phi$ is unital, then:

  1. (3)

    $V$ is an isometry.

  2. (4)

    Under the identification that $H = V(H) \subset K$, $(K, \pi(a))$ is a dilation of $\phi(a)$ for all $a \in A$.

The triple $(\pi, V, K)$ is a Stinespring representation of $\phi$. If $\bracsn{\pi(a)VH|a \in A}$ is dense in $K$, then $(\pi, V, K)$ is minimal[1].

Proof, [Theorem 4.1, Po02]. (a): Let $A \otimes H$ be the algebraic tensor product and define the Hermitian form

\[\dpn{\cdot, \cdot}{\phi}: (A \otimes H) \times (A \otimes H)\to \complex \quad \dpn{a \otimes \xi, b \otimes \eta}{\phi}= \dpn{\phi(b^*a)\xi, \eta}{H}\]

For each $\sum_{j = 1}^{n} a_{j} \otimes \xi_{j} \in A \otimes H$, let $\xi = (\xi_{1}, \cdots, \xi_{n}) \in H^{n}$ and

\[a = \sum_{i, j = 1}^{n} E_{ij}\otimes a_{i}^{*} a_{j} = \braks{\sum_{i = 1}^n E_{1i} \otimes a_{i}}^{*}\braks{\sum_{j = 1}^n E_{1j} \otimes a_j}\ge 0\]

then since $\phi$ is completely positive,

\[\angles{\sum_{j = 1}^n a_j \otimes \xi_j, \sum_{j = 1}^n a_j \otimes \xi_j}_{\phi} = \sum_{i, j = 1}^{n} \dpn{\phi(a_i^*a_j)\xi_j, \xi_i}{H}= \dpn{\phi_n(a)\xi, \xi}{H^n}\ge 0\]

and $\dpn{\cdot, \cdot}{\phi}$ is a pseudo-inner product.

Let $N = \bracsn{u \in A \otimes H| \dpn{u, u}{\phi} = 0}$. For any $u \in N$ and $v \in A \otimes H$, by the Cauchy-Schwarz inequality,

\[|\dpn{u, v}{\phi}| \le \dpn{u, u}{\phi}^{1/2}\dpn{v, v}{\phi}^{1/2}= 0\]

so $N = \bracsn{u \in A \otimes H|\dpn{u, v}{\phi} = 0 \forall v \in A \otimes H}$.

Let $K_{0} = (A \otimes H)/N$ and $K$ be the completion of $K_{0}$ with respect to $\dpn{\cdot, \cdot}{\phi}$, then $K$ is a complex Hilbert space.

(b): Let $V: H \to K_{0}$ be defined by $\xi \mapsto 1_{A} \otimes \xi + N$, then for each $\xi \in H$,

\[\norm{V\xi}_{K_0}^{2} = \dpn{1_A \otimes \xi, 1_A \otimes \xi}{\phi}= \dpn{\phi(1_A)\xi, \xi}{H}\le \norm{\phi(1_A)}_{B(H)}\norm{\xi}_{H}^{2}\]

so $\norm{V}_{L(H; K)}^{2} \le \norm{\phi(1_A)}_{B(H)}$.

(c): For each $a \in A$, let

\[\pi(a): K_{0} \to K_{0} \quad \pi(a)(b \otimes \xi + N) = (ab) \otimes \xi + N\]

then for any $u = \sum_{j = 1}^{n} b_{j} \otimes \xi_{j} \in A \otimes H$,

\begin{align*}\norm{\pi(a)(u+N)}_{K_0}^{2}&= \sum_{i, j = 1}^{n} \dpn{(ab_j) \otimes \xi_j, (ab_i) \otimes \xi_i}{\phi}= \sum_{i, j = 1}^{n} \dpn{\phi(b_i^*a^*ab_j)\xi_j, \xi_i}{H}\\&\le \sum_{i, j = 1}^{n} \norm{a^*a}_{A} \dpn{\phi(b_i^*b_j)\xi_j, \xi_i}{H}= \norm{a}_{A}^{2} \cdot \norm{u + N}_{K_0}^{2}\end{align*}

so $\norm{\pi(a)}_{B(K_0)}\le \norm{a}_{A}$. By the Linear Extension Theorem, $\pi(a)$ extends uniquely to an operator in $B(K)$, and $\pi: A \to B(K)$ is a bounded linear map.

By definition, $\pi$ is a unital homomorphism of Banach algebras. Finally, for any $a \in A$ and $b \otimes \xi, c \otimes \eta \in A \otimes H$,

\begin{align*}\dpn{\pi(a^*)(b \otimes \xi + N), c \otimes \eta + N}{K_0}&= \dpn{(a^*b) \otimes \xi, c \otimes \eta}{\phi}\\&= \dpn{\phi(c^*a^*b)\xi, \eta}{H}= \dpn{b \otimes \xi, (ac) \otimes \eta}{\phi}\\&= \dpn{b \otimes \xi + N, \pi(a)(c \otimes \eta + N)}{K_0}\end{align*}

By linearity, $\dpn{\pi(a^*)u, v}{K_0}= \dpn{u, \pi(a)v}{K_0}$ for all $u, v \in K_{0}$. As $K_{0}$ is dense in $K$, $\pi(a^{*}) = \pi(a)^{*}$, so $\pi$ is a unital *-homomorphism.

(1): For each $\xi \in H$,

\[\norm{V\xi}_{K_0}^{2} = \dpn{1_A \otimes \xi, 1_A \otimes \xi}{\phi}= \dpn{\phi(1_A)\xi, \xi}{H}\]

so

\[\norm{V}_{K_0}^{2} = \sup_{\substack{\xi \in H \\ \norm{\xi}_H \le 1}}\norm{V\xi}_{K_0}^{2} = \sup_{\substack{\xi \in H \\ \norm{\xi}_H \le 1}}\dpn{\phi(1_A)\xi, \xi}{H}= w(\phi(1_{A})) = \norm{\phi(1_A)}_{B(H)}\]

by Proposition 41.6.2.

(2): For each $a \in A$ and $\xi, \eta \in H$,

\[\dpn{V^*\pi(a)V\xi, \eta}{H}= \dpn{\pi(a)V\xi, V\eta}{K}= \dpn{a \otimes \xi, 1_A \otimes \eta}{\phi}= \dpn{\phi(a)\xi, \eta}{H}\]

so $V^{*}\pi(a)V = \phi(a)$.

(3): If $\phi$ is unital, then for each $\xi \in H$,

\[\dpn{V\xi, V\xi}{K}= \dpn{1_A \otimes \xi, 1_A \otimes \xi}{\phi}= \dpn{\phi(1_A)\xi, \xi}{H}= \dpn{\xi, \xi}{H}\]

so $V$ is an isometry.

(4): Suppose that $\phi$ is unital, then $V$ is an isometry by (3). Under the identification that $H = V(H) \subset K$, let $P \in B(K)$ be the orthogonal projection onto $H$, then for each $a \in A$ and $\xi, \eta \in H$,

\[\dpn{\phi(a)\xi, \eta}{H}= \dpn{\pi(a)V\xi, V\eta}{K}= \dpn{\pi(a)P\xi, P\eta}{K}= \dpn{P\pi(a)\xi, \eta}{K}\]

Therefore $\phi(a) = P\pi(a)|_{H}$, and $(K, \pi(a))$ is a dilation of $\phi(a)$.$\square$

Proposition 41.8.6 (Uniqueness of Minimal Stinespring Representations).label Let $A$ be a unital $C^{*}$-algebra, $H$ be a complex Hilbert space, $\phi: A \to B(H)$ be a completely positive map, and $(\pi_{1}, V_{1}, K_{1}), (\pi_{2}, V_{2}, K_{2})$ be minimal Stinespring representations for $\phi$, then there exists a unitary equivalence $U: K_{1} \to K_{2}$ such that $UV_{1} = V_{2}$.

Proof, [Proposition 4.2, Po02]. Suppose there exists a unitary equivalence $U: K_{1} \to K_{2}$ such that $UV_{1} = V_{2}$, then for each $\seqf{a_j}\subset A$ and $\seqf{\xi_j}\subset H$,

\[U\braks{\sum_{j = 1}^n \pi_1(a_j)V_1\xi_j}= \sum_{j = 1}^{n} \pi_{2}(a_{j})UV_{1}\xi_{j} = \sum_{j = 1}^{n} \pi_{2}(a_{j})V_{2}\xi_{j}\]

By minimality, $\pi_{1}(A)V_{1}H$ is dense in $K_{1}$, and $\pi_{2}(A)V_{2}H$ is dense in $K_{2}$. As such, it is sufficient to show that the mapping $\sum_{j = 1}^{n} \pi_{1}(a_{j})V_{1}\xi_{j} \mapsto \sum_{j = 1}^{n} \pi_{2}(a_{j})V_{2}\xi_{j}$ is a well-defined isometry. To this end, note that

\begin{align*}\norm{\sum_{j = 1}^n \pi_1(a_j)V_1\xi_j}_{K_1}^{2}&= \sum_{i, j = 1}^{n} \dpn{\pi_1(a_j)V_1\xi_j, \pi_1(a_i)V_1\xi_i}{K_1}\\&= \sum_{i, j = 1}^{n} \dpn{V_1^*\pi_1(a_i^*a_j)V_1\xi_j, \xi_i}{H}= \sum_{i, j = 1}^{n} \dpn{\phi(a_i^*a_j)\xi_j, \xi_i}{H}\\&= \sum_{i, j = 1}^{n} \dpn{V_2^*\pi_2(a_i^*a_j)V_2\xi_j, \xi_i}{H}\\&= \sum_{i, j = 1}^{n} \dpn{\pi_2(a_j)V_2\xi_j, \pi_2(a_i)V_2\xi_i}{K_2}= \norm{\sum_{j = 1}^n \pi_2(a_j)V_2\xi_j}_{K_2}^{2}\end{align*}

$\square$

Theorem 41.8.7 (Sz.-Nagy’s Dilation Theorem).label Let $H$ be a complex Hilbert space and $T \in B(H)$ with $\norm{T}_{B(H)}\le 1$, then there exists a unitary power dilation $(K, U)$ of $T$ such that:

  1. (1)

    $K$ is the smallest closed reducing subspace for $U$ containing $H$ such that $T^{n} = P_{H}U^{n}|_{H}$[2] for all $n \in \natz$.

  2. (U)

    For any unitary power dilation $(K', U')$ satisfying (1), there exists a unitary $V: K \to K'$ such that $V|_{H}= I_{H}$ and $VUV^{*} = U'$.

Constructive proof without uniqueness, [Theorem 1.1, Po02]. Using Lemma 41.8.4, it is sufficient to show the following statements:

  1. (i)

    For any $T \in B(H)$ with $\norm{T}_{B(H)}\le 1$, there exists an isometric power dilation of $T$.

  2. (ii)

    For any isometry $V \in B(H)$, there exists a unitary power dilation of $V$.

(i): Since $\norm{T}_{B(H)}\le 1$, $1 - T^{*}T \ge 0$, and $(1 - T^{*}T)^{1/2}$ is well-defined. Let $D_{T} = (1 - T^{*}T)^{1/2}$, and

\[V: l^{2}(\natp; H) \to l^{2}(\natp; H) \quad (Vx)_{n} = \begin{cases}Tx_{1}&n = 1 \\ D_{T}x_{1}&n = 2 \\ x_{n-1}&n \ge 3\end{cases}\]

then for each $x \in H$,

\begin{align*}\norm{Tx}_{H}^{2} + \norm{D_Tx}_{H}^{2}&= \dpn{T^*Tx, x}{H}+ \dpn{(1 - T^*T)x, x}{H}\\&= \dpn{x, x}{H}= \norm{x}_{H}^{2}\end{align*}

Thus for each $x = \seq{x_n}\in l^{2}(H)$,

\begin{align*}\norm{Vx}_{l^2(\natp; H)}^{2}&= \norm{Tx_1}_{H}^{2} + \norm{D_Tx_1}_{H}^{2} + \sum_{n = 3}^{\infty} \norm{x_{n-1}}_{H}^{2} \\&= \norm{x_1}_{H}^{2} + \sum_{n = 2}^{\infty} \norm{x_{n}}_{H}^{2} = \norm{x}_{l^2(\natp; H)}^{2}\end{align*}

So $V$ is an isometry. Moreover, $[V^{n}(x, 0, \cdots)]_{1} = T^{n}x$ for all $n \in \natz$. Therefore under the identification that $H = \bracsn{(x, 0, \cdots)|x \in H}\subset l^{2}(\natp; H)$, $V$ is an isometric power dilation of $T$.

(ii): Let $P = I_{H} - VV^{*}$ be the projection onto $V(H)^{\perp}$, and $U \in B(H^{2})$ be defined by

\[U = \begin{bmatrix}V&P \\ 0&V^{*}\end{bmatrix}\]

then

\begin{align*}UU^{*}&= \begin{bmatrix}V&P \\ 0&V^{*}\end{bmatrix} \cdot \begin{bmatrix}V^{*}&0 \\ P&V\end{bmatrix} = \begin{bmatrix}VV^{*} + P&PV \\ V^{*}P&V^{*}V\end{bmatrix} = I_{H^2}\\ U^{*}U&= \begin{bmatrix}V^{*}&0 \\ P&V\end{bmatrix} \cdot \begin{bmatrix}V&P \\ 0&V^{*}\end{bmatrix} = \begin{bmatrix}V^{*}V&V^{*}P \\ PV&VV^{*} + P\end{bmatrix} = I_{H^2}\\\end{align*}

so $U$ is unitary. Moreover, for each $n \in \natp$,

\[U^{n} = \begin{bmatrix}V&P \\ 0&V^{*}\end{bmatrix}^{n} = \begin{bmatrix}V^{n}&\cdots \\ 0&(V^{*})^{n}\end{bmatrix}\]

Under the identification that $H = \bracsn{(x, 0)|x \in H}\subset H^{2}$, let $Q \in B(H^{2})$ be the projection onto the first summand, then $QU^{n}|_{H}= V^{n}$ for all $n \in \natp$. Therefore $U$ is a unitary power dilation of $V$.$\square$

Proof, using Stinespring’s Dilation Theorem, [Theorem 4.3, Po02]. (1): Let $\mathbb{T}= \partial B_{\complex}(0, 1)$. By Theorem 41.7.3, the mapping

\[\phi: \complex[z, \ol z] \to B(H) \quad \phi(p + \ol q) = p(T) + q(T)^{*}\]

extends to a unital positive map $C(\mathbb{T}; \complex) \to B(H)$, which is completely positive by Stinespring’s Theorem.

Let $(\pi, V, K)$ be a minimal Stinespring representation for $\phi$ and $U := \pi(z) \in U(B(K))$, then for each $n \in \natz$,

\[T^{n} = \phi(z^{n}) = P_{H}\pi(z^{n})|_{H} = P_{H}U^{n}|_{H}\]

so $(K, U)$ is a unitary power dilation of $T$.

(U): By the Stone-Weierstrass Theorem, $(\pi, V, K)$ is minimal if and only if the span of $\bigcup_{n \in \integer}U^{n} H$ is dense in $K$, if and only if $U$ admits no non-trivial reducing subspace.

By uniqueness of minimal Stinespring representations, such a dilation is unique.$\square$

Theorem 41.8.8 (Von Neumann’s Inequality).label Let $H$ be a complex Hilbert space and $T \in B(H)$ with $\norm{T}_{B(H)}\le 1$, then for any $p, q \in \complex[x]$,

\[\normn{p(T) + q(T)^*}_{B(H)}\le \sup_{z \in B_\complex(0, 1)}|p(z) + \ol{q(z)}|\]

Proof, [Corollary 1.2, Po02]. By Sz.-Nagy’s Dilation Theorem, $T$ admits a unitary power dilation $(K, U)$. Let $P \in B(K)$ be the orthogonal projection onto $H$, then by Lemma 41.8.4,

\[p(T) + q(T)^{*} = P[p(U) + q(U)^{*}]|_{H}\]

As $U$ is unitary, the continuous functional calculus shows that $\norm{p(U) + q(U)^*}_{B(K)}\le \sup_{z \in B_\complex(0, 1)}|p(z) + \ol{q(z)}|$. Therefore

\[\normn{p(T) + q(T)^*}_{B(H)}\le \sup_{z \in B_\complex(0, 1)}|p(z) + \ol{q(z)}|\]

$\square$

Proof, [Exercise 2.15, Po02]. Assume without loss of generality that $\norm{T}_{B(H)}< 1$. Let

\[P(T; t) = \frac{1}{1 - e^{-it}T}+ \frac{1}{1 - e^{it}T^{*}}- 1\]

Let $p, q \in \complex[z]$ and $\gamma: [0, 2\pi] \to \complex$ be defined by $\gamma(t) = e^{it}$, then by the holomorphic functional calculus,

\[p(T) = \frac{1}{2\pi i}\int_{\gamma} \frac{p(\lambda)}{\lambda - T}d\lambda\]

For each $n \in \natz$, since $\gamma([0, 2\pi]) \subset \partial B_{\complex}(0, 1)$, $\ol{\lambda^n}= \lambda^{-n}$. After passing through a Neumann series,

\begin{align*}\int_{\gamma} \frac{\ol{\lambda^n}}{\lambda - T}d\lambda&= \int_{\gamma} \frac{\lambda^{-n}}{\lambda - T}d\lambda = \int_{\gamma} \frac{\lambda^{-{(n+1)}}}{1 - T/\lambda}d\lambda \\&= \sum_{k = 0}^{\infty} T^{k}\int_{\gamma}\lambda^{-(n+k+1)}d\lambda = \begin{cases}2\pi i & n = 0 \\ 0 & n > 0\end{cases}\end{align*}

By linearity and Cauchy’s integral formula,

\[\frac{1}{2\pi i}\int_{\gamma}\frac{\ol{q(\lambda)}}{\lambda - T}d\lambda = \ol{q(0)}= \frac{1}{2\pi i}\int_{\gamma} \frac{\ol{q(\lambda)}}{\lambda}d\lambda\]

so

\begin{align*}p(T) + \overline{q(0)}&= \frac{1}{2\pi i}\int_{\gamma} \frac{p(\lambda) + \ol{q(\lambda)}}{\lambda - T}d\lambda \\ p(T) + \overline{q(0)}&= \frac{1}{2\pi i}\int_{0}^{2\pi}\frac{ie^{it}[p(e^{it}) + \ol{q(e^{it})}]}{e^{it}- T}dt \\&= \frac{1}{2\pi}\int_{0}^{2\pi}\frac{p(e^{it}) + \ol{q(e^{it})}}{1 - e^{-it}T}dt\end{align*}

Similarly,

\[q(T)^{*} + p(0) = \frac{1}{2\pi}\int_{0}^{2\pi}\frac{p(e^{it}) + \ol{q(e^{it})}}{1 - e^{it}T^{*}}dt\]

Therefore

\begin{align*}p(T) + q(T)^{*} + p(0) + \ol{q(0)}&= \frac{1}{2\pi}\int_{0}^{2\pi}[p(e^{it}) + \ol{q(e^{it})}](P(T; t) + 1)dt \\ p(T) + q(T)^{*}&= \frac{1}{2\pi}\int_{0}^{2\pi}[p(e^{it}) + \ol{q(e^{it})}]P(T; t)dt\end{align*}

Now, for each $t \in [0, 2\pi]$,

\[(1 - e^{-it}T)(1 - e^{it}T^{*}) = 1 - e^{-it}T - e^{it}T^{*} + TT^{*}\]

so by the resolvent identity,

\begin{align*}P(T; t)&= \frac{1}{1 - e^{-it}T}+ \frac{1}{1 - e^{it}T^{*}}- 1 \\&= (1 - e^{-it}T)^{-1}\braks{1 - TT^*}(1 - e^{it}T^{*})^{-1}\ge 0\end{align*}

Let $M = \sup_{z \in \partial B_\complex(0, 1)}|p(z) + \ol{q(z)}|$. For each $\xi, \eta \in H$, let $\dpn{\xi, \eta}{t}= \dpn{P(T; t)\xi, \eta}{H}$, then $\dpn{\cdot, \cdot}{t}$ is a pseudo inner product on $H$. Hence by the Cauchy-Schwarz inequality,

\begin{align*}\dpn{[p(T) + q(T)^*]\xi, \eta}{H}&= \frac{1}{2\pi}\int_{0}^{2\pi}[p(e^{it}) + \ol{q(e^{it})}]\dpn{P(T; t)\xi, \eta}{H}dt \\&= \frac{1}{2\pi}\int_{0}^{2\pi}[p(e^{it}) + \ol{q(e^{it})}]\dpn{\xi, \eta}{t}dt \\ |\dpn{[p(T) + q(T)^*]\xi, \eta}{H}|&\le \frac{1}{2\pi}\int_{0}^{2\pi}M\dpn{\xi, \xi}{t}^{1/2}\dpn{\eta, \eta}{t}^{1/2}dt\\&\le M\braks{\frac{1}{2\pi}\int_0^{2\pi}\dpn{\xi, \xi}{t} dt}^{1/2}\braks{\frac{1}{2\pi}\int_0^{2\pi}\dpn{\eta, \eta}{t} dt}^{1/2}\end{align*}

Since $I_{H} = \frac{1}{2\pi}\int_{0}^{2\pi}P(T; t)dt$,

\[\frac{1}{2\pi}\int_{0}^{2\pi}\dpn{\xi, \xi}{t}dt = \frac{1}{2\pi}\int_{0}^{2\pi}\dpn{P(T; t)\xi, \xi}{H}dt = \norm{\xi}_{H}^{2}\]

Therefore $\dpn{[p(T) + q(T)^*]\xi, \eta}{H}\le M\norm{\xi}_{H}\norm{\eta}_{H}$, and $\norm{[p(T) + q(T)^*]}_{B(H)}\le M$.$\square$

Theorem 41.8.9 (Von Neumann’s Inequality for Matrices).label Let $n \in \natp$, $T \in M_{n}(\complex)$ with $\norm{T}_{M_n(\complex)}\le 1$, and $U, V, S \in M_{n}(\complex)$ such that:

  1. (a)

    $U, V$ are unitary.

  2. (b)

    $S = \text{diag}(s_{1}, \cdots, s_{n})$, where $\seqf{s_j}\subset [0, 1]$.

  3. (c)

    $T = USV$.

Let $B = \ol{B_\complex(0, 1)}^{n} \subset \complex^{n}$ and

\[T: B \to M_{n}(\complex) \quad T(z_{1}, \cdots, z_{n}) = U\text{diag}(z_{1}, \cdots, z_{n})V\]

and $p \in \complex[z]$, then:

  1. (1)

    For each $\xi, \eta \in \complex^{n}$, let

    \[f_{\xi, \eta}: B \to \complex \quad (z_{1}, \cdots, z_{n}) \mapsto \dpn{p(T(z_1, \cdots, z_n))\xi, \eta}{\complex^n}\]

    then $f_{\xi, \eta}$ achieves its maximum modulus at a point $(z_{1}, \cdots, z_{n}) \in \complex^{n}$ such that $|z_{j}| = 1$ for all $1 \le j \le n$.

  2. (2)

    For any $(z_{1}, \cdots, z_{n}) \in \complex^{n}$, if

    \[\normn{p(T(z_1, \cdots, z_n))}_{M_n(\complex)}= \max_{w \in \overline{B}}\normn{p(T(w_1, \cdots, w_n))}_{M_n(\complex)}\]

    then $T(z_{1}, \cdots, z_{n})$ is unitary.

  3. (3)

    $\norm{p(T)}_{M_n(\complex)}\le \max_{W \in U(M_n(\complex))}\norm{p(W)}_{M_n(\complex)}$.

  4. (4)

    $\norm{p(T)}_{M_n(\complex)}\le \norm{p}_{H^\infty(D)}$.

Proof, [Exercise 2.16, Po02], attributed to Wermer. (1): By the Maximum Modulus Principle.

(2): Let $(z_{1}, \cdots, z_{n}) \in \complex^{n}$ such that $\normn{p(T(z_1, \cdots, z_n))}_{M_n(\complex)}$ is maximal. By compactness of the unit sphere in $\complex^{n}$, there exists $\xi, \eta \in \partial B_{\complex^n}(0, 1)$ such that

\[\normn{p(T(z_1, \cdots, z_n))}_{M_n(\complex)}= \dpn{p(T(z_1, \cdots, z_n))\xi, \eta}{\complex^n}= f_{\xi, \eta}(z_{1}, \cdots, z_{n})\]

By maximality, $f_{\xi, \eta}$ achieves its maximum modulus at $(z_{1}, \cdots, z_{n})$, so (1) implies that $|z_{j}| = 1$ for all $1 \le j \le n$ and $\text{diag}(z_{1}, \cdots, z_{n})$ is unitary. Therefore $T(z_{1}, \cdots, z_{n}) = U\text{diag}(z_{1}, \cdots, z_{n})V$ is also unitary.

(3): Since $(s_{1}, \cdots, s_{n}) \in B$, (2) implies that

\begin{align*}\norm{p(T)}_{M_n(\complex)}&\le \max_{z \in B}\norm{p(T(z_1, \cdots, z_n))}_{M_n(\complex)}\\&\le \max_{W \in U(M_n(\complex))}\norm{p(W)}_{M_n(\complex)}\end{align*}

(4): For any $W \in U(M_{n}(\complex))$, the continuous functional calculus implies that

\[\norm{p(W)}_{M_n(\complex)}\le \max_{\lambda \in \sigma_{M_n(\complex)}(W)}|p(\lambda)| \le \max_{\lambda \in \partial B_\complex(0, 1)}|p(\lambda)| = \norm{p}_{H^\infty(D)}\]

As the above holds for all unitary matrices,

\[\norm{p(T)}_{M_n(\complex)}\le \max_{W \in U(M_n(\complex))}\norm{p(W)}_{M_n(\complex)}\le \norm{p}_{H^\infty(D)}\]

by (3).$\square$

  1. $K$ can always be trimmed so that $(\pi, V, K)$ is minimal.keyboard_return
  2. Where $P_{H}$ is the orthogonal projection onto $H$.keyboard_return

Post a Comment

Name:Email:
Please enter the tag of the current page (1HX) to post the comment.
Tag: