6
$\begingroup$

Let $A$ be a $N\times N$-matrix with elements $$ a_{ii}=1 \quad\text{and}\quad a_{ij} = \frac{1}{ij} \quad\text{for}~ i\neq j. $$ Then $A$ is positive-definite, as can be easily seen from $$ x^T A x = \sum_i x_i^2 + \sum_{i \neq j} \frac{x_i x_j}{ij} \geq \sum_i \frac{x_i^2}{i^2} + \sum_{i \neq j} \frac{x_i x_j}{ij} = \left(\sum_i \frac{x_i}{i}\right)^2 \geq 0. $$

Assume now that $A$ is a real symmetric $N\times N$-matrix with elements $$ \tag{1} a_{ii}=1 \quad\text{and}\quad |a_{ij}| \leq \frac{1}{ij} \quad\text{for}~ i\neq j. $$

Is it possible to show that $A$ is also positive-definite (or positive-semidefinite)?


Here I posted a refinement of this question by assuming $0 \leq a_{ij} \leq \frac{1}{ij}$ for $i \neq j$ in $(1)$.

$\endgroup$
4
  • 1
    $\begingroup$ it may be better to think of this in terms of congruence transforms -- with diagonal $D := \text{diag}(1,2,...,n)$ then your original $A$ is congruent to $DAD = D^2 - I +\mathbf {11}^T$ and for $\big \Vert \mathbf x \big \Vert_2 = 1$ we have $\mathbf x^T DAD \mathbf x = \mathbf x^T \big(D^2 - I +\mathbf {11}^T\big)\mathbf x \geq \mathbf x^T\big(D^2 - I\big)\mathbf x \geq 0$ where the second lower bound is met with equality iff $\mathbf x \propto \mathbf e_1$, in which case the first bound is strict... this shows positive definiteness. $\endgroup$ Commented Feb 25, 2020 at 23:40
  • $\begingroup$ Could you say something about the origin of this question ? $\endgroup$
    – Jean Marie
    Commented Feb 27, 2020 at 10:48
  • 1
    $\begingroup$ @JeanMarie Positive-definiteness of such matrix appears as a sufficient condition for the completeness in $L^2(0,1)$ of a certain perturbation of the system of square wave functions, and follows from the Paley-Wiener stability theorem (see, e.g., Theorem 9.2 in Singer's "Bases in Banach Spaces I"). The values $\frac{1}{ij}$ are scalar products of corresponding square waves. Basically, if certain perturbations of the system of square waves are sufficiently small, then the elements $a_{ij}$ have to be close to $\frac{1}{ij}$, and, presumably, the completeness property should be kept. $\endgroup$ Commented Feb 28, 2020 at 22:02
  • 1
    $\begingroup$ Thank you for your answer. I posted some years ago a question here about structured matrices having a certain similarity with yours but with complex eigenvalues. I have never had satisfactory answers. If, by chance, you have an idea... $\endgroup$
    – Jean Marie
    Commented Feb 28, 2020 at 22:24

2 Answers 2

5
$\begingroup$

It's false.

Choose $N=20$ and, for every $i\not= j$, $A_{i,j}=\dfrac{-1}{ij}$. $A$ admits a negative eigenvalue $\approx -0.04$.

It suffices to consider your first matrix $A$ -denoted by $A0$ (with $A0_{i,j}=1/(ij)$)-.

After, increase $N$ until $\rho(A0)> 2$.

$\endgroup$
5
  • $\begingroup$ Thank you! Indeed, you are right. But I've just realized that in the actual matrix which I work with (it comes from approximation theory) all elements are positive. Could you advice me, please, should I post a new question about it, or there is also a simple counterexample or a way to get a result? $\endgroup$ Commented Feb 26, 2020 at 7:38
  • 1
    $\begingroup$ If the $(a_{i,j})$ are $> 0$, then I think it's right, but I haven't really thought about this new situation. It's better that you post a new question. $\endgroup$
    – user91684
    Commented Feb 26, 2020 at 8:52
  • 1
    $\begingroup$ 1) I realize that simulations don't give always good hints. 2) In fact, one can start from $N=13$ where the smallest eigenvalue is $-0.004$. $\endgroup$
    – Jean Marie
    Commented Feb 26, 2020 at 8:58
  • 1
    $\begingroup$ @JeanMarie Yes, the simulations work poorly when counterexamples are in "corners". $\endgroup$
    – user91684
    Commented Feb 26, 2020 at 9:01
  • 1
    $\begingroup$ An odd fact : on an experimental, basis, vector $V$ with coordinates $V_k=k/(1+k^2)$ is a very good approximation of the eigenvector associated to the first (negative) eigenvalue. $\endgroup$
    – Jean Marie
    Commented Feb 26, 2020 at 9:40
4
$\begingroup$

Precise computations dealing with the solution by loup blanc:

Let $$A_N=\begin{pmatrix} 1&-1/2&-1/3&\cdots&-1/N\\ -1/2&1&-1/6&\cdots&-1/(2N)\\ -1/3&-1/6&1&\cdots&-1/(3N)\\ \cdots&\cdots&\cdots&\cdots&\cdots\\ -1/N&-1/(2N)&-1/(3N)&\cdots&1\\ \end{pmatrix}$$

(diagonal elements equal to $1$, off-diagonal elements $(A_N)_{ij}=-\dfrac{1}{ij}$.)

We are going to obtain a formula for $\det(A_N)$ (formula (2)) showing a certain value $N_0$ exists such that

$$\text{for all} \ N \geq N_0, \ \ \det(A_{N})<0$$

Let us use the following formula for the determinant of an invertible matrix $M$ and (column) vectors $U,V \in \mathbb{R^n}$

$$\det(M+UV^T)=\det(M)(1+V^TM^{-1}U)\tag{1}$$

(matrix-determinant lemma)

If we take :

$$\begin{cases}M&=&diag(1+1/1^2,1+1/2^2,1+1/3^2, \cdots 1+1/N^2)\\ U&=&(1,1/2,1/3,\cdots 1/N)^T\\ V&=&-U\end{cases},$$

we have :

$$A_N=M+UV^T.$$

Therefore, using (1) :

$$\det(A_N)=\prod_{k=1}^N \left(1+\dfrac{1}{k^2}\right) \times \left(1-\sum_{k=1}^N \dfrac{1}{k} \times \dfrac{1}{1+\tfrac{1}{k^2}} \times \dfrac{1}{k}\right)$$

$$\det(A_N)=\prod_{k=1}^N \left(1+\dfrac{1}{k^2}\right) \times \left(1-\sum_{k=1}^N \dfrac{1}{k^2+1}\right)\tag{2}$$

But the following series is convergent with sum :

$$\sum_{k=1}^{+\infty} \dfrac{1}{k^2+1}=\dfrac12\left(\pi \coth \pi -1 \right)\approx 1.07667...> 1$$

(see a proof here or there).

Therefore, the content of the second parenthesis in (2) becomes negative beyond a certain $N_0$ as expected ; thus $\det(A_N)<0 $ for $N \geq N_0$ : $A_N$ cannot be semi-definite positive for $N \geq N_0$.

This value $N_0$ happens to be $13$ as "forecasted" by numerical attempts.

Remarks :

1) The first part of the determinant above tends, when $N \to \infty$ to the following convergent infinite product (see this) :

$$\prod_{k=1}^{\infty} \left(1+\dfrac{1}{k^2}\right)=\dfrac{\text{sinh} \ \pi}{\pi}\approx 3.676078.$$

As a consequence, for very large values of $N$,

$$\det(A_N) \approx \dfrac{1}{2}\left(\dfrac{1}{\pi}(3 \sinh \pi)- \cosh \pi\right) = -0.28186..$$

2) It is instructive to plot the eigenvalues of $A_N$ (which are all real because $A_N$ is symmetrical). See figure (horizontal axis : values of $N$ from $N=1$ to $N=200$). Different aspects can be underlined, in particular the fact that there is a unique negative eigenvalue ($\approx -0.10720$ for $N$ sufficiently large), a huge cluster just above $1$, and separate values at $\approx 1.0858226$, $\approx 1.176735$, and $\approx 1.593347$.

These curves are either increasing or decreasing as a consequence of Cauchy interlacing theorem for symmetric matrices.

enter image description here

A possible connection :(https://www.sciencedirect.com/science/article/pii/S002437951200198X).

$\endgroup$
2
  • $\begingroup$ @loup blanc I have given some precisions in this answer about the minimal size $A$ must have in order to have $\det(A)<0$ : I find back in this way $N_0=13$... $\endgroup$
    – Jean Marie
    Commented Feb 26, 2020 at 22:56
  • 1
    $\begingroup$ Very elegant approach! $\endgroup$ Commented Feb 28, 2020 at 8:09

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .