2
$\begingroup$

Let $\mathcal M_2$ be the space of real $2\times 2$ matrices and $\mathcal S_2\subset \mathcal M_2$ be its subset consisting of positive semidefinite elements, i.e. $A\in \mathcal S_2$ iff $A$ is symmetric and $u^TAu\ge 0$ for all $u\in\mathbb R^2$. We consider the distance between $A,B\in \mathcal S_2$ given by $$d(A,B):={\rm Tr}(A + B - 2\sqrt{\sqrt{A}B\sqrt{A}}),$$ where ${\rm Tr}$ denotes the trace operator. My question is as follows:

Define the unit ball $B_1 :=\{u=(u_1,u_2)\in \mathbb R^2: u_1^2+u_2^2\le 1\}$. Set $$L(A,B):=\max_{u\in B_1} \Big\{ u^TAu+u^TBu-2\sqrt{u^TAuu^TBu}\Big\}.$$ Does the maximisation problem below admit a finite upper bound? $$\sup_{A, B\in \mathcal S_2}\frac{d(A,B)}{L(A,B)}<\infty?$$

PS : It is straightforward to observe $d(A,B)=d(B,A)$ and $d(cA,cB)=cd(A,B)$ for $c\ge 0$ (where the same relations hold for $L$).

$\endgroup$
9
  • $\begingroup$ The arguments seem to be very calculative, while elementary calculus yield very heavy expression to determine whether the upper bound is finite or not. I wonder whether there exist advanced results/tools in matrix theory that help $\endgroup$ Commented May 22, 2024 at 19:03
  • $\begingroup$ What do you mean by $\sqrt{AB}$? $\endgroup$ Commented May 22, 2024 at 19:33
  • $\begingroup$ @IosifPinelis This denotes the square root of the matrix $AB$ that is positive semidefinite by assumption. More precisely, ${\rm Tr}(\sqrt{AB})$ is uniquely defined. $\endgroup$ Commented May 22, 2024 at 19:37
  • 1
    $\begingroup$ $AB$ is not symmetric or positive semidefinite in general. So, what is $\sqrt{AB}$? $\endgroup$ Commented May 22, 2024 at 20:04
  • 1
    $\begingroup$ @GerryMyerson On my behalf, the display is correct for Safari but not for Google Chrome. Thanks anyway for fixing this issue :) $\endgroup$ Commented May 23, 2024 at 8:29

1 Answer 1

7
$\begingroup$

The answer is "yes" and it is quite a nice linear algebra problem, but let me restate it first in a less intimidating way. We'll deal with $\mathbb R^n$ for any finite $n$.

The first thing I will do is to replace $A$ with $A^2$ and $B$ with $B^2$. Next, notice that $\newcommand{\tr}{\operatorname{Tr}}$ $\tr\sqrt{AB^2A}=\sup_{u,v}\sum_k\langle BAu_k,v_k\rangle=\sup_{u,v}\sum_k\langle Au_k,Bv_k\rangle$ where $u,v$ are two orthonormal bases in $\mathbb R^n$. Indeed, if $C$ is any linear operator in $\mathbb R^n$, then writing $C=RD$ where $R$ is orthogonal and $D$ is self-adjoint positive semi-definite, we get $\sqrt{C^*C}=D$. Also the expression $\sup_{u,v}\sum_k\langle Cu_k,v_k\rangle$ is invariant under multiplications of $C$ by orthogonal operators from either side, so the supremum for $C$ is the same as for $D$. However, for $D$, we have $$ \sum_k\langle Du_k,v_k\rangle\le\frac 12\left[\sum_k\langle Du_k,u_k\rangle+\sum_k\langle Dv_k,v_k\rangle\right]=\tr D $$ with equality when $u=v$.

Since $\tr A^2=\sum_k |Au_k|^2$ and $\tr B^2=\sum_k |Bu_k|^2$ regardless of the choice of $u,v$, we see that the LHS is just $$ \inf_{u,v}\sum_k|Au_k-Bv_k|^2 $$ Let $R$ be the orthogonal mapping defined by $Ru_k=v_k$. Then the infimum can be rewritten as $$ \inf_R\sum_k|(A-BR)u_k|^2=\inf_R\|A-BR\|_{HS}^2=\inf_R\|A-RB\|_{HS}^2\,. $$ (the last equality is due to the fact that $\|A-BR\|_{HS}=\|A-R^*B\|_{HS}$).

The RHS in this notation is merely $\sup_x(|Ax|-|Bx|)^2$ where $x$ runs over the unit sphere. Thus we arrive to the question if for two operators $A$ and $B$, the only reason why $\left||Ax|-|Bx|\right|<\varepsilon$ for all unit vectors $x$ is that $\|A-RB\|\le K\varepsilon$ for some orthogonal $R$ with some $K=K(n)\in[1,+\infty)$ (obviously, if the latter condition holds with $K=1$, then the former one is satisfied as well). What norm to use doesn't matter in the finite dimension, so I'll switch to the operator norm now.

What I suspect is that $K$ should be independent of the dimension in this setting. However, I can do it only with $K(n)$ growing exponentially in $n$ at this moment, which is rather dismal. Still, it solves the problem as posted.

We shall proceed by induction on $n$. The base $n=1$ is trivial with $K(1)=1$.

For the step $n-1\to n$, we can always assume that $A$ is self-adjoint PSD by multiplying it by an orthogonal operator on the left. Also, since the inequality scales correctly, we can normalize to $\|A\|=1$. If $\varepsilon>\frac 13$, we can take $K=7$, so I'll assume that $0<\varepsilon\le\frac 13$.

Then $\|B\|\le\frac 43$. Denote by $e$ the top unit eigenvector of $A$, so $Ae=e$. Then $\mu=|Be|\in[\frac 23,\frac 43]$. Let $R'$ be the rotation such that $B'e=R'Be=\mu e$. Consider now a unit vector $y\in e^\perp$ and put $x=(\cos t)e+(\sin t)y$. On the one hand, we have $$ ||Ax|^2-|B'x|^2|=||Ax|^2-|Bx|^2|=||Ax|-|Bx||(|Ax|+|Bx|)\le \frac 73\varepsilon\,. $$ On the other hand, using that $Ae^\perp\subset e^\perp$, the left hand side can be written as $$ |J_1\cos^2 t+J_2\sin^2 t-\mu\langle e,B'y\rangle\sin 2t| $$ where $J_1$ and $J_2$ is some junk independent of $t$. Since $\sin 2t$ is orthogonal to $\sin^2 t$ and $\cos^2 t$ in $L^2(-\pi,\pi)$, we conclude that $$ \mu|\langle e,B'y\rangle|\le \frac{14}{3}\varepsilon\,, $$ so, since $\mu\ge \frac 23$, we get $$ |\langle e,B'y\rangle|\le 7\varepsilon\,. $$ Now consider the operator $B''=B'-S$ where $Sx=\langle B'P_{e^\perp}x,e\rangle e$. Note that $\|S\|\le 7\varepsilon$, so $||Ax|-|B''x||\le 8\varepsilon$ for all unit $x$. Also $B''e=B'e=\mu e$ and $e^\perp$ is invariant under both $A$ and $B''$. Thus, applying the induction assumption, we can find $R''$ that rotates only in $e^\perp$ (i.e., $R''e=e$) such that $\|A-R''B''\|_{e^\perp\to e^\perp}\le 8K(n-1)\varepsilon$. Together with $|(A-R''B'')e|=|(A-B')e|\le\varepsilon$, it yields $$ \|A-R''B''\|\le 8K(n-1)\varepsilon $$ and, finally, $$ \|A-R''R'B\|=\|A-R''B'\|\le \|A-R''B''\|+\|S\|\le [8K(n-1)+7]\varepsilon\,, $$ i.e., we can take $K(n)=8K(n-1)+7$.

As I said, this bound (growing as $8^n$) is ridiculously large and it would be nice to figure out what really happens here. Any ideas?

$\endgroup$
5
  • 1
    $\begingroup$ Thank you so much for this answer and I find your idea amazing! I always struggled when falling in the analytic reasoning (treating this problem as a traditional optimisation problem) $\endgroup$ Commented May 24, 2024 at 7:08
  • 1
    $\begingroup$ I realise that you have answered a question which is quite analytic, and you showed your algebra skills/techniques in this post $\endgroup$ Commented May 24, 2024 at 7:09
  • $\begingroup$ Hi @fedja I asked a similar question in mathoverflow.net/questions/472985/… I didn't succeed in extending your analysis to this case. Do you think the same argument works? $\endgroup$ Commented Jun 10, 2024 at 14:22
  • $\begingroup$ Dear @fedja I think I missed one step in ur proof: Around 7 lines above the end of your proof, you claim that "$|(A-R''B'')e|=|(A-B')e|\le \varepsilon$". However, one has $|(A-B')e|=|e-\mu e|=|1-\mu|\le 1/3$. As it is only known $0<\varepsilon\le 1/3$, how could we deduce $|(A-B')e|\le \varepsilon$? $\endgroup$ Commented May 6 at 18:07
  • $\begingroup$ @Fawen90 We actually know (from the definition of $\mu$ and our assumption about $A$ and $B$) that $|1-\mu|=||Ae|-|Be||\le \varepsilon$. Most of the proof doesn't use the full strength of this inequality and relaxes it to $1/3$ on the RHS, but at the place you referred to we need the full version. $\endgroup$ Commented May 30 at 19:30

You must log in to answer this question.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.