17
$\begingroup$

Let $N$ be a nilpotent matrix in $M_n({\mathbb R})$, such that $(I+N)^2$ is “positive definite” (but not necessarily symmetric) in the sense that $<X,(I+N)^2X>$ is positive for any nonzero $X\in{{\mathbb R}^n}$ (here $<.,.>$ denotes the usual scalar product on ${\mathbb R}^n$). Is it true that $I+N$ must always be “positive definite” also ?

UPDATE : To clarify the question, I know about http://mathworld.wolfram.com/PositiveDefiniteMatrix.html which makes many promising related statements, but without any proof. This webpage has a list of references at the bottom, but I’ve checked them one by one and was unable to find that special result.

I think this question is a rather special case of the properties of the "unique positive square root function" described in the link. I'm hoping for an “elementary” proof.

$\endgroup$
  • $\begingroup$ Indeed -- in fact, this result holds for any non-singular (which $(I+N)^2$ is, since otherwise our nilpotent $N$ has a non-zero eigenvalue!), positive definite matrix M. $\endgroup$ – FH93 Feb 10 '14 at 8:58
  • $\begingroup$ @FH93 I’m not sure you understood the question, as you just said the answer is “YES” without any proof or reference $\endgroup$ – Ewan Delanoy Feb 10 '14 at 10:54
  • $\begingroup$ I don't know whether it is true or not, but it seems to me that the numerical range of such matrices satisfies $W((I+N)^k)\subset W((I+N)^{k-1})$, $k=1,2,\ldots$ (I was looking before for a counterexample). So if this is true, then the statement in the question is a consequence. $\endgroup$ – Algebraic Pavel Feb 10 '14 at 14:22
  • $\begingroup$ @AlgebraicPavel What is it that you denote by $W$ ? The kernel? $\endgroup$ – Ewan Delanoy Feb 10 '14 at 14:41
  • 2
    $\begingroup$ I wonder if the second paragraph in mathworld would not be misplaced; in particular, it is talking about the Choleski theorem that is reserved for $>0$ symmetric matrices. I have the impression that the whole paragaraph concerns only $>0$ symmetric matrices. Anyway, I never heard that "unique positive square root" theorem would be also valid for non-symmetric $> 0 $ matrices. Yet, during several numerical tests, I observed that the numerical range of $I+N$ is included in the numerical range of $(I+N)^2$ (when $(I+N)^2>0$).... $\endgroup$ – loup blanc Nov 25 '17 at 0:00
7
+100
$\begingroup$

I believe $I+N$ must indeed be "positive definite".

Apply induction on the dimension. We are given a nilpotent matrix $N'$ which by Schur decomposition we can assume is strictly upper triangular. As a block matrix, $N'=\begin{pmatrix}N&x\\0&0\end{pmatrix}$ with $N$ a square strictly upper triangular matrix and $x$ a column vector. By assumption $(I+N')^2+(I+N')^{2t}$ is SPD (SPD=symmetric positive definite; superscript $2t$ means square and transpose). We wish to show that $(I+N')+(I+N')^t$ is SPD.

Let $U=I+N$ and note $$\begin{pmatrix}U&x\\0&1\end{pmatrix}^2+\begin{pmatrix}U&x\\0&1\end{pmatrix}^{2t}=\begin{pmatrix}U^2+U^{2t}&(I+U)x\\x^t(I+U)^t&2\end{pmatrix}.$$

By the Schur decomposition characterization of SPD we have by assumption:

  • $U^2+U^{2t}$ is SPD
  • $2-x^t(I+U)^t(U^2+U^{2t})^{-1}(I+U)x>0$

And we need to show:

  • $U+U^t$ is SPD
  • $2-x^t(U+U^t)^{-1}x>0.$

$U+U^t$ is SPD by induction. The latter inequality would follow from $$x^t(I+U)^t(U^2+U^{2t})^{-1}(I+U)x\geq x^t(U+U^t)^{-1}x$$ Letting $y=(I+U)x,$ this is $$y^t(U^2+U^{2t})^{-1}y\geq y^t(I+U^t)^{-1}(U+U^t)^{-1}(I+U)^{-1}y$$

This would follow from this result and symmetric positive definiteness of $$(I+U)(U+U^t)(I+U^t)-(U^2+U^{2t})=U+U^t+2UU^t+U(U+U^t)U^t$$ Since $U+U^t$ is SPD, so is $U(U+U^t)U^t,$ and $UU^t$ is certainly SPD. So we are done.


See loup blanc's answer for a generalization of the above argument from unipotent matrices to all matrices with positive spectrum.


Here is a reference with a nice proof: Uniqueness of matrix square roots and an application by Charles R. Johnson, Kazuyoshi Okubo, Robert Reams, Theorem 7. It uses the following theorem:

Theorem (Lyapunov). Let $A\in\mathbb C^{n\times n}$ (not necessarily Hermitian) and let $X\in\mathbb C^{n\times n}$ be Hermitian. If the eigenvalues of $A$ all have positive real part and $AX+XA^*$ is positive definite, then $X$ is positive definite.

In particular if $A$ has eigenvalues with positive real part and $A(A+A^*)+(A+A^*)A^*=A^2+(A^*)^2+2AA^*$ is positive definite, then $A+A^*$ is positive definite.

Proof 1 (sketch) following Horn and Johnson's Topics in Matrix Analysis:

Suppose not. We can take a kind of Jordan normal form, but making any above-diagonal $1$'s arbitrarily small, so $S^{-1}AS+(S^{-1}AS)^*$ is positive definite for some non-singular $S.$ Setting $G=SS^*$ we find that $AG+GA^*$ is positive definite. For $0\leq \theta\leq 1$ define $X_{\theta}=\theta G+(1-\theta)X.$ Note:

  • $X_0=X$ is not positive definite
  • $X_1=G$ is positive definite
  • $AX_\theta+X_\theta A^*$ is a convex combination of positive definite matrices so must be positive definite.

All the matrices $X_\theta$ have real eigenvalues, and by continuity some $X_\theta$ must have $0$ as an eigenvalue: $X_\theta v=0$ for some non-zero $v.$ This implies $v^*(AX_\theta+X_\theta A^*)v=0,$ contradicting positive definiteness of $AX_\theta+X_\theta A^*.$

Proof 2.

Consider the function defined by $f(W)=\int_0^\infty e^{-tA}We^{-tA^*}dt.$ We can compute

\begin{align*} W&=-\frac d{d\tau}\Bigr|_{\tau=0}\int_{\tau}^\infty e^{-tA}We^{-tA^*}dt\\ &=-\frac{d}{d\tau}\Bigr|_{\tau=0}\int_0^\infty e^{-\tau A}e^{-tA}We^{-tA^*}e^{-\tau A^*}dt\\ &=Af(W)+f(W)A^*. \end{align*}

This means $f$ is a left inverse of the map $X\mapsto AX+XA^*.$ Since $f$ is an $\mathbb R$-linear map from the space of Hermitian matrices to itself, and has a left inverse, it must be an isomorphism. So the solution $X$ of $AX+XA^*=W$ is unique. And if $W$ is positive definite, then so is $e^{-tA}W^{1/2}W^{1/2}e^{-tA^*},$ so $f(W)$ is positive definite by construction. This proves that if $AX+XA^*$ is positive definite then so is $X.$

$\endgroup$
  • $\begingroup$ Beautiful proof. I'll wait before awarding the bounty, but I doubt that anyone will come with something better, given the difficulty of the problem. $\endgroup$ – Ewan Delanoy Nov 27 '17 at 15:17
  • $\begingroup$ @EwanDelanoy: I have added a reference which you might be interested in, and paraphrased their proof $\endgroup$ – Dap Nov 28 '17 at 5:52
3
$\begingroup$

A very partial answer.

Proposition 1. let $A$ be a P.D. matrix. Then there is a unique $B$ s.t. $B^2=A$ and every eigenvalue $\lambda$ of $B$ satisfies $Re(\lambda)>0$; moreover, if $A$ admits a P.D. square root, then necessarily it's $B$.

Proof. The key point is: if $U\in M_n$ is P.D., then every eigenvalue $\mu$ of $U$ satisfies $Re(\mu)>0$ (Beware, the converse is false!)

In particular, our $A$ has no $<0$ eigenvalues and, therefore, admits a unique square root $B$ s.t. every eigenvalue $\lambda$ of $B$ satisfies $Re(\lambda)>0$ (cf. Higham, functions of matrices). Thus $B$ is the only candidate that can be P.D.

Remark. That does not imply (despite Mathworld's article) that $A$ admits a P.D. square root.

EDIT 1. @Dap did a very pretty proof. I had thought of making such a recurrence, but I was sure that it would not work; which just shows that, in mathematics, you have to believe!

Moreover, using Dap's proof (mutatis mutandis), we can prove the following improvement

Proposition 2. Let $A\in M_n(\mathbb{R})$ be a P.D. matrix that satisfies $spectrum(A)\subset (0,+\infty)$. Then its principal square root (cf. Proposition 1.) is P.D.

Proof. Note that $B$ (the principal square root of $A$) has $>0$ eigenvalues and that it is triangularizable over $\mathbb{R}$ with a change of orthonormal basis.

Let $N'=\begin{pmatrix}N&x\\0&\alpha\end{pmatrix}$ (where $\alpha>0$) be the matrix $B$ after triangularization. We follow the Dap's proof.

We know that $N^2+N^{2T}>0,2\alpha^2-x^T(N+\alpha I)^T(N^2+N^{2T})^{-1}(N+\alpha I)x>0$ and we want to show that

$N+N^T>0,2\alpha-x^T(N+N^T)^{-1}x>0$.

It suffices to show that

$\Delta=x^T(N+\alpha I)^T(N^2+N^{2T})^{-1}(N+\alpha I)x- \alpha(x^T(N+N^T)^{-1}x)$ is non-negative.

We find $\Delta=N(N+N^T)N^T+2\alpha NN^T+\alpha^2(N+N^T)$ and we are done.

Remark 1. It remains to study the case when a P.D. matrix $A$ has non-real eigenvalues with positive real part.

Remark 2. Of course, Dap deserves the bounty.

EDIT 2. I just read the article ([1] by Johnson and all) cited by @Dap. Ewan will be happy; the result given by mathworld is true (what surprises me).

If $A\in M_n(\mathbb{C})$, let $F(A)=\{x^*Ax;||x||=1,x\in \mathbb{C}^n\}$ be its numerical range or its field of values. Note that $A$ is P.D. iff $F(A)\subset \{z;Re(z)>0\}$ iff $A+A^*$ is H.P.D..

[1] Theorem 7 (Kato,Masser, Neumann). If $A\in M_n(\mathbb{C})$ is s.t. $F(A)\cap (-\infty,0]=\emptyset$ (that is the case when $A$ is P.D.), then its principal square root is P.D..

[1] Corollary 8 (Johnson and all). If $A\in M_n(\mathbb{R})$ is s.t. $F(A)\cap (-\infty,0]=\emptyset$ (that is the case when $A$ is P.D. in the sense: for every $x\in \mathbb{R}^n\setminus\{0\}$, $x^TAx>0$), then its principal square root (which is a real matrix) is P.D.

$\endgroup$
  • $\begingroup$ The obvious next question is : since the converse is false in general, is it true when $U$ is of the form $I+N$ where $N$ is nilpotent. Alas, it doesn't work : when $N=\left(\begin{array}{cc} 0 & 3 \\ 0 & 0\end{array}\right)$, then $I+N$'s only eigenvalue is $1$ but it is not P.D. Too bad ... $\endgroup$ – Ewan Delanoy Nov 26 '17 at 5:23

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.