Entrywise absolute value matrix and second largest eigenvalue.

The name of the pictureThe name of the pictureThe name of the pictureClash Royale CLAN TAG#URR8PPP











up vote
1
down vote

favorite
1












Suppose $A=(a_ij)$ is a Hermitian $ntimes n$ matrix such that the symmetric matrix $vert Arvert=(lvert a_ijrvert)$ has a largest eigenvalue $lambda_1(lvert Arvert)=1$, and all other eigenvalues satisfy $$frac12ge lambda_2(lvert Arvert)gedotsge lambda_n(vert Arvert)ge-frac12.$$



Let $x$ be the normalised eigenvector of $A$ corresponding to eigenvalue $lambda_1(A)$. Then
$$lambda_1(A)=langle x,Axrangle = lvertlangle x,Axranglervert le langle lvert xrvert,lvert Arvertlvert xrvertranglele lambda_1(lvert Arvert),$$
but what can be said about $lambda_2(A)$? For example is it true that for sufficiently small $epsilon>0$ there exists a small $delta=delta(epsilon)>0$ such that
$$lvert lambda_1(A)-1rvert<deltaquadimpliesquad lambda_2(A)lefrac12+epsilon?$$
If yes, how can $delta(epsilon)$ be chosen?







share|cite|improve this question






















  • Why is it clear that $lambda_1(A)lelambda_1(|A|)$?
    – amsmath
    Aug 28 at 15:24











  • @amsmath See my edit.
    – Julian
    Aug 28 at 15:36










  • I think you want $|lambda_1(A)-1| < delta implies lambda_2(A) le frac12+epsilon$. Otherwise just take $delta = 1$.
    – Robert Israel
    Aug 28 at 15:55










  • @RobertIsrael Yes, thanks for spotting this typo.
    – Julian
    Aug 28 at 15:57










  • @Julian Why is $langle x,Axrangle = |langle x,Axrangle|$? This is not the case if $A$ is negative definite. EDIT: Ok, but then we trivially have $lambda_1(A)lelambda_1(|A|)$. Thanks for answering my question.
    – amsmath
    Aug 28 at 18:08















up vote
1
down vote

favorite
1












Suppose $A=(a_ij)$ is a Hermitian $ntimes n$ matrix such that the symmetric matrix $vert Arvert=(lvert a_ijrvert)$ has a largest eigenvalue $lambda_1(lvert Arvert)=1$, and all other eigenvalues satisfy $$frac12ge lambda_2(lvert Arvert)gedotsge lambda_n(vert Arvert)ge-frac12.$$



Let $x$ be the normalised eigenvector of $A$ corresponding to eigenvalue $lambda_1(A)$. Then
$$lambda_1(A)=langle x,Axrangle = lvertlangle x,Axranglervert le langle lvert xrvert,lvert Arvertlvert xrvertranglele lambda_1(lvert Arvert),$$
but what can be said about $lambda_2(A)$? For example is it true that for sufficiently small $epsilon>0$ there exists a small $delta=delta(epsilon)>0$ such that
$$lvert lambda_1(A)-1rvert<deltaquadimpliesquad lambda_2(A)lefrac12+epsilon?$$
If yes, how can $delta(epsilon)$ be chosen?







share|cite|improve this question






















  • Why is it clear that $lambda_1(A)lelambda_1(|A|)$?
    – amsmath
    Aug 28 at 15:24











  • @amsmath See my edit.
    – Julian
    Aug 28 at 15:36










  • I think you want $|lambda_1(A)-1| < delta implies lambda_2(A) le frac12+epsilon$. Otherwise just take $delta = 1$.
    – Robert Israel
    Aug 28 at 15:55










  • @RobertIsrael Yes, thanks for spotting this typo.
    – Julian
    Aug 28 at 15:57










  • @Julian Why is $langle x,Axrangle = |langle x,Axrangle|$? This is not the case if $A$ is negative definite. EDIT: Ok, but then we trivially have $lambda_1(A)lelambda_1(|A|)$. Thanks for answering my question.
    – amsmath
    Aug 28 at 18:08













up vote
1
down vote

favorite
1









up vote
1
down vote

favorite
1






1





Suppose $A=(a_ij)$ is a Hermitian $ntimes n$ matrix such that the symmetric matrix $vert Arvert=(lvert a_ijrvert)$ has a largest eigenvalue $lambda_1(lvert Arvert)=1$, and all other eigenvalues satisfy $$frac12ge lambda_2(lvert Arvert)gedotsge lambda_n(vert Arvert)ge-frac12.$$



Let $x$ be the normalised eigenvector of $A$ corresponding to eigenvalue $lambda_1(A)$. Then
$$lambda_1(A)=langle x,Axrangle = lvertlangle x,Axranglervert le langle lvert xrvert,lvert Arvertlvert xrvertranglele lambda_1(lvert Arvert),$$
but what can be said about $lambda_2(A)$? For example is it true that for sufficiently small $epsilon>0$ there exists a small $delta=delta(epsilon)>0$ such that
$$lvert lambda_1(A)-1rvert<deltaquadimpliesquad lambda_2(A)lefrac12+epsilon?$$
If yes, how can $delta(epsilon)$ be chosen?







share|cite|improve this question














Suppose $A=(a_ij)$ is a Hermitian $ntimes n$ matrix such that the symmetric matrix $vert Arvert=(lvert a_ijrvert)$ has a largest eigenvalue $lambda_1(lvert Arvert)=1$, and all other eigenvalues satisfy $$frac12ge lambda_2(lvert Arvert)gedotsge lambda_n(vert Arvert)ge-frac12.$$



Let $x$ be the normalised eigenvector of $A$ corresponding to eigenvalue $lambda_1(A)$. Then
$$lambda_1(A)=langle x,Axrangle = lvertlangle x,Axranglervert le langle lvert xrvert,lvert Arvertlvert xrvertranglele lambda_1(lvert Arvert),$$
but what can be said about $lambda_2(A)$? For example is it true that for sufficiently small $epsilon>0$ there exists a small $delta=delta(epsilon)>0$ such that
$$lvert lambda_1(A)-1rvert<deltaquadimpliesquad lambda_2(A)lefrac12+epsilon?$$
If yes, how can $delta(epsilon)$ be chosen?









share|cite|improve this question













share|cite|improve this question




share|cite|improve this question








edited Aug 28 at 15:57

























asked Aug 28 at 15:20









Julian

1,026925




1,026925











  • Why is it clear that $lambda_1(A)lelambda_1(|A|)$?
    – amsmath
    Aug 28 at 15:24











  • @amsmath See my edit.
    – Julian
    Aug 28 at 15:36










  • I think you want $|lambda_1(A)-1| < delta implies lambda_2(A) le frac12+epsilon$. Otherwise just take $delta = 1$.
    – Robert Israel
    Aug 28 at 15:55










  • @RobertIsrael Yes, thanks for spotting this typo.
    – Julian
    Aug 28 at 15:57










  • @Julian Why is $langle x,Axrangle = |langle x,Axrangle|$? This is not the case if $A$ is negative definite. EDIT: Ok, but then we trivially have $lambda_1(A)lelambda_1(|A|)$. Thanks for answering my question.
    – amsmath
    Aug 28 at 18:08

















  • Why is it clear that $lambda_1(A)lelambda_1(|A|)$?
    – amsmath
    Aug 28 at 15:24











  • @amsmath See my edit.
    – Julian
    Aug 28 at 15:36










  • I think you want $|lambda_1(A)-1| < delta implies lambda_2(A) le frac12+epsilon$. Otherwise just take $delta = 1$.
    – Robert Israel
    Aug 28 at 15:55










  • @RobertIsrael Yes, thanks for spotting this typo.
    – Julian
    Aug 28 at 15:57










  • @Julian Why is $langle x,Axrangle = |langle x,Axrangle|$? This is not the case if $A$ is negative definite. EDIT: Ok, but then we trivially have $lambda_1(A)lelambda_1(|A|)$. Thanks for answering my question.
    – amsmath
    Aug 28 at 18:08
















Why is it clear that $lambda_1(A)lelambda_1(|A|)$?
– amsmath
Aug 28 at 15:24





Why is it clear that $lambda_1(A)lelambda_1(|A|)$?
– amsmath
Aug 28 at 15:24













@amsmath See my edit.
– Julian
Aug 28 at 15:36




@amsmath See my edit.
– Julian
Aug 28 at 15:36












I think you want $|lambda_1(A)-1| < delta implies lambda_2(A) le frac12+epsilon$. Otherwise just take $delta = 1$.
– Robert Israel
Aug 28 at 15:55




I think you want $|lambda_1(A)-1| < delta implies lambda_2(A) le frac12+epsilon$. Otherwise just take $delta = 1$.
– Robert Israel
Aug 28 at 15:55












@RobertIsrael Yes, thanks for spotting this typo.
– Julian
Aug 28 at 15:57




@RobertIsrael Yes, thanks for spotting this typo.
– Julian
Aug 28 at 15:57












@Julian Why is $langle x,Axrangle = |langle x,Axrangle|$? This is not the case if $A$ is negative definite. EDIT: Ok, but then we trivially have $lambda_1(A)lelambda_1(|A|)$. Thanks for answering my question.
– amsmath
Aug 28 at 18:08





@Julian Why is $langle x,Axrangle = |langle x,Axrangle|$? This is not the case if $A$ is negative definite. EDIT: Ok, but then we trivially have $lambda_1(A)lelambda_1(|A|)$. Thanks for answering my question.
– amsmath
Aug 28 at 18:08











1 Answer
1






active

oldest

votes

















up vote
0
down vote













The answer to your question is yes.



First of all, consider the case $lambda_1(A) = 1$.
Note that $A_ij = 0$ if $x_i ne 0$ and $x_j = 0$: otherwise you could increase $langle y, |A| y rangle/langle y, yrangle$ by taking $y = |x| + t e_j$ for some small $t$ (the numerator is $1 + c t + O(t^2)$ for some $c > 0$, the denominator $1 + O(t^2)$).
Thus (reordering the rows and columns if necessary) if some $x_i = 0$ the matrix $A$ has block form $pmatrixB & 0cr 0 & Ccr$, and its eigenvalues are the union of the eigenvalues of $B$ and of $C$, and similarly for $|A|$ with $|B|$ and $|C|$; $1$ is an eigenvalue of $B$ and of $|B|$ while all eigenvalues of $|C|$ are at most $1/2$. Then all eigenvalues of $C$ have the same bound. Thus we can restrict our attention to $B$, i.e. we may assume all $x_i ne 0$.



Let $U$ be a diagonal matrix with diagonal entries $u_ii = x_i/|x_i|$.
We have $langle x, A x rangle = langle |x|, |A| |x| rangle = 1$ which
implies all the terms $overlinex_i A_ij x_j ge 0$, i.e. $|x_i| overlineu_ii A_ij u_jj |x_j| ge 0$, so $overlineu_ii A_ij u_jj ge 0$. This says $|A| = U^* A U$. But then $|A|$ and $A$ have the same eigenvalues, so $|lambda_2(A)| le 1/2$.



Finally, note that the set of Hermitian matrices satisfying your condition is compact.
If the answer were no, then there would be $epsilon > 0$ and a sequence of such matrices $A_k$ with
$lambda_1(A_k) to 1$ and $lambda_2(A_k) > 1/2 + epsilon$. Taking a limit point, we get $A$ satisfying the condition with $lambda_1(A) = 1$ and $lambda_2(A) ge 1/2 + epsilon$, which as we have seen is impossible.






share|cite|improve this answer




















  • Thanks a lot. It seems that the soft compactness argument allows for $delta$ to depend on the dimension $n$? I was rather asking about $delta$ only depending on $epsilon$. I am wondering whether any quantitative estimate could be extracted from the compactness argument?
    – Julian
    Aug 28 at 17:37











Your Answer




StackExchange.ifUsing("editor", function ()
return StackExchange.using("mathjaxEditing", function ()
StackExchange.MarkdownEditor.creationCallbacks.add(function (editor, postfix)
StackExchange.mathjaxEditing.prepareWmdForMathJax(editor, postfix, [["$", "$"], ["\\(","\\)"]]);
);
);
, "mathjax-editing");

StackExchange.ready(function()
var channelOptions =
tags: "".split(" "),
id: "69"
;
initTagRenderer("".split(" "), "".split(" "), channelOptions);

StackExchange.using("externalEditor", function()
// Have to fire editor after snippets, if snippets enabled
if (StackExchange.settings.snippets.snippetsEnabled)
StackExchange.using("snippets", function()
createEditor();
);

else
createEditor();

);

function createEditor()
StackExchange.prepareEditor(
heartbeatType: 'answer',
convertImagesToLinks: true,
noModals: false,
showLowRepImageUploadWarning: true,
reputationToPostImages: 10,
bindNavPrevention: true,
postfix: "",
noCode: true, onDemand: true,
discardSelector: ".discard-answer"
,immediatelyShowMarkdownHelp:true
);



);













 

draft saved


draft discarded


















StackExchange.ready(
function ()
StackExchange.openid.initPostLogin('.new-post-login', 'https%3a%2f%2fmath.stackexchange.com%2fquestions%2f2897390%2fentrywise-absolute-value-matrix-and-second-largest-eigenvalue%23new-answer', 'question_page');

);

Post as a guest






























1 Answer
1






active

oldest

votes








1 Answer
1






active

oldest

votes









active

oldest

votes






active

oldest

votes








up vote
0
down vote













The answer to your question is yes.



First of all, consider the case $lambda_1(A) = 1$.
Note that $A_ij = 0$ if $x_i ne 0$ and $x_j = 0$: otherwise you could increase $langle y, |A| y rangle/langle y, yrangle$ by taking $y = |x| + t e_j$ for some small $t$ (the numerator is $1 + c t + O(t^2)$ for some $c > 0$, the denominator $1 + O(t^2)$).
Thus (reordering the rows and columns if necessary) if some $x_i = 0$ the matrix $A$ has block form $pmatrixB & 0cr 0 & Ccr$, and its eigenvalues are the union of the eigenvalues of $B$ and of $C$, and similarly for $|A|$ with $|B|$ and $|C|$; $1$ is an eigenvalue of $B$ and of $|B|$ while all eigenvalues of $|C|$ are at most $1/2$. Then all eigenvalues of $C$ have the same bound. Thus we can restrict our attention to $B$, i.e. we may assume all $x_i ne 0$.



Let $U$ be a diagonal matrix with diagonal entries $u_ii = x_i/|x_i|$.
We have $langle x, A x rangle = langle |x|, |A| |x| rangle = 1$ which
implies all the terms $overlinex_i A_ij x_j ge 0$, i.e. $|x_i| overlineu_ii A_ij u_jj |x_j| ge 0$, so $overlineu_ii A_ij u_jj ge 0$. This says $|A| = U^* A U$. But then $|A|$ and $A$ have the same eigenvalues, so $|lambda_2(A)| le 1/2$.



Finally, note that the set of Hermitian matrices satisfying your condition is compact.
If the answer were no, then there would be $epsilon > 0$ and a sequence of such matrices $A_k$ with
$lambda_1(A_k) to 1$ and $lambda_2(A_k) > 1/2 + epsilon$. Taking a limit point, we get $A$ satisfying the condition with $lambda_1(A) = 1$ and $lambda_2(A) ge 1/2 + epsilon$, which as we have seen is impossible.






share|cite|improve this answer




















  • Thanks a lot. It seems that the soft compactness argument allows for $delta$ to depend on the dimension $n$? I was rather asking about $delta$ only depending on $epsilon$. I am wondering whether any quantitative estimate could be extracted from the compactness argument?
    – Julian
    Aug 28 at 17:37















up vote
0
down vote













The answer to your question is yes.



First of all, consider the case $lambda_1(A) = 1$.
Note that $A_ij = 0$ if $x_i ne 0$ and $x_j = 0$: otherwise you could increase $langle y, |A| y rangle/langle y, yrangle$ by taking $y = |x| + t e_j$ for some small $t$ (the numerator is $1 + c t + O(t^2)$ for some $c > 0$, the denominator $1 + O(t^2)$).
Thus (reordering the rows and columns if necessary) if some $x_i = 0$ the matrix $A$ has block form $pmatrixB & 0cr 0 & Ccr$, and its eigenvalues are the union of the eigenvalues of $B$ and of $C$, and similarly for $|A|$ with $|B|$ and $|C|$; $1$ is an eigenvalue of $B$ and of $|B|$ while all eigenvalues of $|C|$ are at most $1/2$. Then all eigenvalues of $C$ have the same bound. Thus we can restrict our attention to $B$, i.e. we may assume all $x_i ne 0$.



Let $U$ be a diagonal matrix with diagonal entries $u_ii = x_i/|x_i|$.
We have $langle x, A x rangle = langle |x|, |A| |x| rangle = 1$ which
implies all the terms $overlinex_i A_ij x_j ge 0$, i.e. $|x_i| overlineu_ii A_ij u_jj |x_j| ge 0$, so $overlineu_ii A_ij u_jj ge 0$. This says $|A| = U^* A U$. But then $|A|$ and $A$ have the same eigenvalues, so $|lambda_2(A)| le 1/2$.



Finally, note that the set of Hermitian matrices satisfying your condition is compact.
If the answer were no, then there would be $epsilon > 0$ and a sequence of such matrices $A_k$ with
$lambda_1(A_k) to 1$ and $lambda_2(A_k) > 1/2 + epsilon$. Taking a limit point, we get $A$ satisfying the condition with $lambda_1(A) = 1$ and $lambda_2(A) ge 1/2 + epsilon$, which as we have seen is impossible.






share|cite|improve this answer




















  • Thanks a lot. It seems that the soft compactness argument allows for $delta$ to depend on the dimension $n$? I was rather asking about $delta$ only depending on $epsilon$. I am wondering whether any quantitative estimate could be extracted from the compactness argument?
    – Julian
    Aug 28 at 17:37













up vote
0
down vote










up vote
0
down vote









The answer to your question is yes.



First of all, consider the case $lambda_1(A) = 1$.
Note that $A_ij = 0$ if $x_i ne 0$ and $x_j = 0$: otherwise you could increase $langle y, |A| y rangle/langle y, yrangle$ by taking $y = |x| + t e_j$ for some small $t$ (the numerator is $1 + c t + O(t^2)$ for some $c > 0$, the denominator $1 + O(t^2)$).
Thus (reordering the rows and columns if necessary) if some $x_i = 0$ the matrix $A$ has block form $pmatrixB & 0cr 0 & Ccr$, and its eigenvalues are the union of the eigenvalues of $B$ and of $C$, and similarly for $|A|$ with $|B|$ and $|C|$; $1$ is an eigenvalue of $B$ and of $|B|$ while all eigenvalues of $|C|$ are at most $1/2$. Then all eigenvalues of $C$ have the same bound. Thus we can restrict our attention to $B$, i.e. we may assume all $x_i ne 0$.



Let $U$ be a diagonal matrix with diagonal entries $u_ii = x_i/|x_i|$.
We have $langle x, A x rangle = langle |x|, |A| |x| rangle = 1$ which
implies all the terms $overlinex_i A_ij x_j ge 0$, i.e. $|x_i| overlineu_ii A_ij u_jj |x_j| ge 0$, so $overlineu_ii A_ij u_jj ge 0$. This says $|A| = U^* A U$. But then $|A|$ and $A$ have the same eigenvalues, so $|lambda_2(A)| le 1/2$.



Finally, note that the set of Hermitian matrices satisfying your condition is compact.
If the answer were no, then there would be $epsilon > 0$ and a sequence of such matrices $A_k$ with
$lambda_1(A_k) to 1$ and $lambda_2(A_k) > 1/2 + epsilon$. Taking a limit point, we get $A$ satisfying the condition with $lambda_1(A) = 1$ and $lambda_2(A) ge 1/2 + epsilon$, which as we have seen is impossible.






share|cite|improve this answer












The answer to your question is yes.



First of all, consider the case $lambda_1(A) = 1$.
Note that $A_ij = 0$ if $x_i ne 0$ and $x_j = 0$: otherwise you could increase $langle y, |A| y rangle/langle y, yrangle$ by taking $y = |x| + t e_j$ for some small $t$ (the numerator is $1 + c t + O(t^2)$ for some $c > 0$, the denominator $1 + O(t^2)$).
Thus (reordering the rows and columns if necessary) if some $x_i = 0$ the matrix $A$ has block form $pmatrixB & 0cr 0 & Ccr$, and its eigenvalues are the union of the eigenvalues of $B$ and of $C$, and similarly for $|A|$ with $|B|$ and $|C|$; $1$ is an eigenvalue of $B$ and of $|B|$ while all eigenvalues of $|C|$ are at most $1/2$. Then all eigenvalues of $C$ have the same bound. Thus we can restrict our attention to $B$, i.e. we may assume all $x_i ne 0$.



Let $U$ be a diagonal matrix with diagonal entries $u_ii = x_i/|x_i|$.
We have $langle x, A x rangle = langle |x|, |A| |x| rangle = 1$ which
implies all the terms $overlinex_i A_ij x_j ge 0$, i.e. $|x_i| overlineu_ii A_ij u_jj |x_j| ge 0$, so $overlineu_ii A_ij u_jj ge 0$. This says $|A| = U^* A U$. But then $|A|$ and $A$ have the same eigenvalues, so $|lambda_2(A)| le 1/2$.



Finally, note that the set of Hermitian matrices satisfying your condition is compact.
If the answer were no, then there would be $epsilon > 0$ and a sequence of such matrices $A_k$ with
$lambda_1(A_k) to 1$ and $lambda_2(A_k) > 1/2 + epsilon$. Taking a limit point, we get $A$ satisfying the condition with $lambda_1(A) = 1$ and $lambda_2(A) ge 1/2 + epsilon$, which as we have seen is impossible.







share|cite|improve this answer












share|cite|improve this answer



share|cite|improve this answer










answered Aug 28 at 17:02









Robert Israel

306k22201443




306k22201443











  • Thanks a lot. It seems that the soft compactness argument allows for $delta$ to depend on the dimension $n$? I was rather asking about $delta$ only depending on $epsilon$. I am wondering whether any quantitative estimate could be extracted from the compactness argument?
    – Julian
    Aug 28 at 17:37

















  • Thanks a lot. It seems that the soft compactness argument allows for $delta$ to depend on the dimension $n$? I was rather asking about $delta$ only depending on $epsilon$. I am wondering whether any quantitative estimate could be extracted from the compactness argument?
    – Julian
    Aug 28 at 17:37
















Thanks a lot. It seems that the soft compactness argument allows for $delta$ to depend on the dimension $n$? I was rather asking about $delta$ only depending on $epsilon$. I am wondering whether any quantitative estimate could be extracted from the compactness argument?
– Julian
Aug 28 at 17:37





Thanks a lot. It seems that the soft compactness argument allows for $delta$ to depend on the dimension $n$? I was rather asking about $delta$ only depending on $epsilon$. I am wondering whether any quantitative estimate could be extracted from the compactness argument?
– Julian
Aug 28 at 17:37


















 

draft saved


draft discarded















































 


draft saved


draft discarded














StackExchange.ready(
function ()
StackExchange.openid.initPostLogin('.new-post-login', 'https%3a%2f%2fmath.stackexchange.com%2fquestions%2f2897390%2fentrywise-absolute-value-matrix-and-second-largest-eigenvalue%23new-answer', 'question_page');

);

Post as a guest













































































這個網誌中的熱門文章

How to combine Bézier curves to a surface?

Why am i infinitely getting the same tweet with the Twitter Search API?

Is there any way to eliminate the singular point to solve this integral by hand or by approximations?