1
$\begingroup$

Let $\mathcal{L}(E)$ be the algebra of all bounded linear operators on a complex Hilbert space $E$. On $\mathcal{L}(E)^2$, we have two equivalent norms: \begin{eqnarray*} N_1(A,B) &=&\sup\left\{\sqrt{\|Ax\|^2+\|Bx\|^2},\;x\in E,\;\|x\|=1\;\right\}, \end{eqnarray*} and $$N_2(A,B)=\sqrt{\|A\|^2+\|B\|^2}.$$

Clearly, $N_1(A,B)\leq N_2(A,B)$.

In general $N_1\neq N_2$. I want to find sufficient conditions for equality.

$\endgroup$
6
  • $\begingroup$ When does $\sup (f + g)$ equal $\sup f + \sup g$? $\endgroup$
    – tsnao
    Jan 8 at 11:39
  • $\begingroup$ Neither of them is homogenous. Did you forget square roots? $\endgroup$ Jan 8 at 12:13
  • $\begingroup$ @JochenWengenroth Yes I'm sorry. I corrected them $\endgroup$
    – Student
    Jan 8 at 13:11
  • $\begingroup$ Here is a sufficient condition: A=B=0. $\endgroup$
    – Yemon Choi
    Jan 9 at 0:18
  • $\begingroup$ The operators $A$ and $B$ are not zero $\endgroup$
    – Student
    Jan 9 at 4:29

1 Answer 1

3
$\begingroup$

$\newcommand\la\lambda\newcommand\ep\varepsilon\newcommand\ip[2]{\langle #1,#2\rangle}\newcommand\Span{\operatorname{span}}$As noted in a comment by Yemon Choi, the question is not well posed.

Apparently, the OP wanted to have a good, nontrivial sufficient condition, preferably close to necessity. Such a sufficient condition will be provided here, and the sufficient condition will also be necessary when $E$ is finite dimensional.

Let \begin{equation} U:=A^*A=\int_{[0,\|A\|^2]}\la\,dP^A_\la\quad\text{and}\quad V:=B^*B=\int_{[0,\|B\|^2]}\la\,dP^B_\la \tag{0}\label{0} \end{equation} be the spectral decompositions of the self-adjoint bounded linear operators $A^*A$ and $B^*B$. (Note that the families $(P^A_\la)$ and $(P^B_\la)$ of orthoprojectors are uniquely determined by condition \eqref{0}.)

Let $$Q^A_\ep:=\int_{[\|A\|^2-\ep,\|A\|^2]}dP^A_\la, \quad Q^B_\ep:=\int_{[\|B\|^2-\ep,\|B\|^2]}dP^B_\la,$$ $$E^A_\ep:=Q^A_\ep E,\quad E^B_\ep:=Q^B_\ep E. $$

Then the condition \begin{equation} E^A_\ep\cap E^B_\ep\ne\{0\}\quad\forall\ep>0 \tag{1}\label{1} \end{equation} is sufficient for $N_1(A,B)=N_2(A,B)$.


Indeed, if \eqref{1} holds, then for each real $\ep>0$ there is a unit vector $$u_\ep\in E^A_\ep\cap E^B_\ep,$$ so that $Q^A_\ep u_\ep=u_\ep=Q^B_\ep u_\ep$. So, $$\begin{aligned} \|Au_\ep\|^2&=\ip{A^*Au_\ep}{u_\ep} \\ &=\int_{[0,\|A\|^2]}\la\,d\ip{P^A_\la u_\ep}{u_\ep} \\ &=\int_{[0,\|A\|^2]}\la\,d\ip{P^A_\la Q^A_\ep u_\ep}{Q^A_\ep u_\ep} \\ &=\int_{[\|A\|^2-\ep,\|A\|^2]}\la\,d\ip{P^A_\la Q^A_\ep u_\ep}{Q^A_\ep u_\ep} \\ &\ge(\|A\|^2-\ep)\int_{[\|A\|^2-\ep,\|A\|^2]}d\ip{P^A_\la Q^A_\ep u_\ep}{Q^A_\ep u_\ep} \\ &=(\|A\|^2-\ep)\ip{Q^A_\ep u_\ep}{Q^A_\ep u_\ep} \\ &=(\|A\|^2-\ep)\ip{u_\ep}{u_\ep} =\|A\|^2-\ep. \end{aligned}$$ Similarly, $\|Bu_\ep\|^2\ge\|B\|^2-\ep$. So, $$N_1(A,B)^2\ge\|Au_\ep\|^2+\|Bu_\ep\|^2 \ge\|A\|^2+\|B\|^2-2\ep=N_2(A,B)^2-2\ep$$ for any real $\ep>0$. Thus, $N_1(A,B)\ge N_2(A,B)\ge N_1(A,B)$. $\quad\Box$


Remark 1: Condition \eqref{1} is necessary for $N_1(A,B)=N_2(A,B)$ when $E$ is finite dimensional. Indeed, suppose that $N_1(A,B)=N_2(A,B)$. Then for some sequence $(x_n)$ on the unit sphere $S$ in $E$ we have \begin{equation} \|A\|^2+\|B\|^2\ge\|Ax_n\|^2+\|Bx_n\|^2\to\|A\|^2+\|B\|^2. \end{equation} It follows that $\|Ax_n\|^2\to\|A\|^2$ and $\|Bx_n\|^2\to\|B\|^2$. The unit sphere $S$ in the finite-dimensional Hilbert space $E$ is compact. So, passing to a subsequence, without loss of generality we have $x_n\to x$ for some $x\in S$. So, $\|Ax\|^2=\|A\|^2$ and $\|Bx\|^2=\|B\|^2$, which implies \eqref{1}. $\quad\Box$

Remark 2: Condition \eqref{1} is not in general necessary for $N_1(A,B)=N_2(A,B)$ when $E$ is infinite dimensional. E.g., let $E:=\ell^2$, with the standard basis $(e_1,e_2,\dots)$.

Let $U$ be the positive-semidefinite self-adjoint linear operator on $E$ whose matrix (in the standard basis) is the block-diagonal matrix with the $2\times2$ diagonal blocks \begin{equation} D_j^U:=\frac j{j+1}\,P_j^U,\quad\text{where}\quad P_j^U:=\begin{bmatrix}1&0\\0&0 \end{bmatrix} \end{equation} for integers $j\ge1$. The factor $\frac j{j+1}$, strictly increasing to $1$ in $j$, was introduced to force gliding to $\infty$ in search of a (nonexistent) maximizer $x\in S$ of $\ip{Ux}x$.

Similarly, let $V$ be the positive-semidefinite self-adjoint linear operator on $E$ whose matrix (in the standard basis) is the block-diagonal matrix with the $2\times2$ diagonal blocks \begin{equation} D_j^V:=\frac j{j+1}\,P_j^V,\quad\text{where}\quad P_j^V:=\frac1{j^2+1}\,\begin{bmatrix}j^2&j\\j&1 \end{bmatrix} \end{equation} for integers $j\ge1$. Note that $P_j^U$ and $P_j^V$ are orthoprojector matrices of rank $1$ each, with the respective column spaces $\Span\Big\{\begin{bmatrix}1\\0 \end{bmatrix}\Big\}$ and $\Span\Big\{\begin{bmatrix}j\\1 \end{bmatrix}\Big\}$.

In accordance with \eqref{0}, let $A:=\sqrt{U}$ and $B:=\sqrt{V}$. Then $\|A\|^2=\|U\|=1$ and $\|B\|^2=\|V\|=1$. Moreover, letting an integer $i$ go to $\infty$, we get $$N_1(A,B)\ge\|Ae_{2i}\|^2+\|Be_{2i}\|^2=\ip{Ue_{2i}}{e_{2i}}+\ip{Ve_{2i}}{e_{2i}} \\ \to1+1=\|A\|^2+\|B\|^2=N_2(A,B),$$ which implies that $N_1(A,B)=N_2(A,B)$.

However, for any integer $j\ge1$ and any $\ep\in(\frac1{j+2},\frac1{j+1}]$, \begin{equation} E_\ep^A=\Span\{e_{2i}\colon i\ge j\}\quad\text{and}\quad E_\ep^B=\Span\{e_{2i}+\tfrac1{2i+1}\,e_{2i+1}\colon i\ge j\}, \end{equation} so that \eqref{1} fails to hold. $\quad\Box$


It appears that the following condition is necessary and sufficient for $N_1(A,B)=N_2(A,B)$:

There exist sequences $(a_n)$ and $(b_n)$ such that $a_n-b_n\to0$, $a_n\in S\cap E_{1/n}^A$ for all $n$, and $b_n\in S\cap E_{1/n}^B$ for all $n$.

This conjecture is probably easy to prove, but ....

$\endgroup$
4
  • 1
    $\begingroup$ I think this should also be a necessary condition, isn’t it? $\endgroup$ Jan 8 at 20:41
  • $\begingroup$ You seem to have lost the square on lambda in the first formula. Or you should specify the projectors' definitios. $\endgroup$
    – tsnao
    Jan 9 at 0:23
  • $\begingroup$ @PietroMajer : This condition is of course necessary if $E$ is finite dimensional. However, it is now shown that it is not necessary in general. $\endgroup$ Jan 9 at 2:25
  • $\begingroup$ @tsnao : I do not think anything is lost here. However, it is now added "that the families $(P^A_\lambda)$ and $(P^B_\lambda)$ of orthoprojectors are uniquely determined by condition (0)". $\endgroup$ Jan 9 at 2:29

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.