12
$\begingroup$

Suppose we have an infinite-dimensional real vector $y=(y_1,...)$. Suppose we have an infinite-dimensional real matrix $C=(c_{ij})$, $i,j\in\mathbb{N}$. Let $C^k$ be a submatrix of $C$, $C^k=(c_{ij})_{i,j=1,k}$ and $y_k$ a subvector of $y$, $y^k=(y_1,...,y_k)$. Define

$$\theta^k=C_k^{-1}y^k$$

My question (which probably is too general) is what conditions should $C$ and $y$ satisfy so that pointwise limits

$$\lim_{k\to\infty}\theta^k_i$$

exist? I got a feeling that this could be easily solved by applying theory of linear operators, but I cannot figure out how to reformulate the problem.

To make this question less general we can assume that $y\in\ell_2$ and $C_k$ is symmetric positive-definite matrix for each $k$.

This question is related to this one I've asked on mathoverflow.

Update @fedja below produces a possible sketch of a proof. It requires though that $\|C_k^{-1}\|$ is bounded sequence (take the matrix norm $\|\|_2$). If we suppose that $C$ is a linear operator in $\ell_2$, does the property that each submatrix $C_k$ is positive-definite ensures that the sequence $\|C_k^{-1}\|$ is bounded?

This question can be further rephrased the following way. For symmetric positive-definite matrix $A$ denote its minimal and maximal eigen values by $\lambda_{\min}$ and $\lambda_{\max}$. Then $\|A\|_2=\lambda_{\max}$ and $\|A^{-1}\|_2=\lambda_{\min}^{-1}$. With this in mind the previous question is identical whether $\lambda_{min}(C_k)$ is bounded away from zero.

Improve this question
$\endgroup$
  • $\begingroup$ Note that not only does the number of equations go to infinity, you also change already existing equations themselves in the process (by adding more variables). $\endgroup$ – Greg Graviton Jul 4 '11 at 7:11
  • $\begingroup$ I am not sure, if this will help you in any way, but Theorem 3.18 in the book [Hanson Yakovlev] Operator Theory for Electromagnetics is about something similar, but it deals with convergence in $\ell_2$, not pointwise convergence. books.google.com/… $\endgroup$ – Martin Sleziak Jul 7 '11 at 6:38
  • $\begingroup$ @mpiktas: I suppose my answer was not exactly what you were hoping for, as you offered a bounty after my answer was there. (I myself wrote that it is closer to a comment than to an answer.) That's why I decided to add a new bounty - perhaps this will attract someones attention and you'll get a better answer. $\endgroup$ – Martin Sleziak Jul 19 '11 at 7:57
  • $\begingroup$ @Martin, thanks. Your answer gave useful additional information, so I gave the bounty to you in order not to waste it. Thanks again for your help! $\endgroup$ – mpiktas Jul 19 '11 at 8:01
  • 1
    $\begingroup$ There is quite a lot of literature devoted to the question how and to what extent a bounded operator $C$ on a Banach or Hilbert space inherits properties of an approximating sequence $(C_n)_n$ of operators with finite-dimensional range. One can study questions like yours in the abstract framework of Banach- and $C^*$-algebras, see for instance Hagen, Roch, Silbermann $C^*$-algebras and numerical analysis. $\endgroup$ – lvb Jul 19 '11 at 10:32
3
+50
$\begingroup$

I'm not sure what exactly you are looking for but if $\|C\|<+\infty$, $\sup_k \|C_k^{-1}\|<+\infty$, and $y\in\ell^2\cap C\ell^2$, then the convergence holds and, moreover, $\theta_k$ converge to the (unique) solution $\theta$ of $C\theta=y$ in $\ell^2$. The reason is very simple. If $P_k$ is the orthogonal projection to the first $k$ coordinates, then $C_k$ is the non-trivial part of $P_kCP_k$ so if $\theta_k$ is extended by zeroes, we have $P_kCP_k\theta_k=P_ky$. On the other hand, $|P_kCP_k\theta-P_ky|\le \|C\|\cdot|P_k\theta-\theta|$. Thus, $|\theta_k-P_k\theta|\le\|C_k^{-1}\|\cdot\|C\|\cdot |P_k\theta-\theta_k|\to 0$ as $k\to\infty$.

Improve this answer
$\endgroup$
  • $\begingroup$ The last term in inequality should be $P_k\theta-\theta$. Good idea, I've already upvoted it. But upon the closer inspection I have a problem with the term $\|C_k^{-1}\|$. Since it depends on $k$ it should be proved that it is bounded. $\endgroup$ – mpiktas Jul 19 '11 at 8:23
2
+50
$\begingroup$

This is closer to a comment than an answer, but it would be too long for a comment. Basically a few thougths that might or might not help. (Just a things this reminded me of, not necessarily helpfull for your problem.)

Let us rewrite your equation as $y^k=C_k\theta^k$. If I consider $y^k$ and $\theta^k$ as infinite sequences (I add zeroes), then I have

$$z^k=C\theta^k,$$

where the vector $z^k$ has the same values as $y^k$ on the first $k$ coordinates.

If we additionaly assume that $y^k\to y$ (which implies $z^k\to y$) and $\theta^k\to\theta$ pointwise and that the operator corresponding to $C$ is in some sense continuous, then we get

$$y=C\theta.$$

Hence a necessary condition is that $y$ is in the range of $C$.

My guess is that this condition should be sufficient for the continuity of $C$:

$$\|C\| = \sup_n \sum_k |c_{nk}| < \infty$$

I am not sure that this really helps but the last equation $$y=C\theta$$ is related to summability theory. In fact, matrix methods studied in summability theory are defined in the way that you transform a sequence using a given matrix and then take a limit.

For overview of some known results you can consult e.g. Boos: Classical and modern methods in summability, you can find also some notes here: http://thales.doa.fmph.uniba.sk/sleziak/texty/rozne/trf/boos/

From a functional analytic viewpoint: Morrison's Functional analysis

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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