On complex eigenvalues of a real nonsymmetric matrix
Andy Wathen · 2025
Read On complex eigenvalues of a real nonsymmetric matrix by Andy Wathen (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-18501.txt.
Introduction
Andy Wathen, in a solo paper from 2025, studies complex eigenvalues of real matrices that are not symmetric. Eigenvalues are the scaling factors along a matrix's special directions, and for real nonsymmetric matrices they can come in complex pairs.
Wathen considers writing such a matrix as a product of real symmetric matrices, a factorization with uses in computation and theory. He proves that the number of complex eigenvalues restricts which factorizations are possible. The result draws a clean line between spectral data, which is easy to compute, and algebraic structure, which is harder to see. Papers of this kind often find use in numerical linear algebra, where symmetric factors are cheaper and stabler to work with.
Catalog record
| CREATOR | Andy Wathen |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 19 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-18501
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
On complex eigenvalues of a real nonsymmetric matrix Andy Wathen Abstract We consider real non-symmetric matrices and their factorisation as a product of real symmetric matrices. The number of complex eigenvalues of the original matrix reveals restrictions on such factorisations as we shall prove. keywords Real nonsymmetric matrices, Complex eigenvalues of real matrices, product of real symmetric matrices AMS 15A18, 15A23 Introduction We all know that real square nonsymmetric matrices can have complex eigenvalues, but how many? Less well-known seems to be the fact that every real square matrix can be expressed as the product of two real symmetric matrices; such a factorisation can often be done in many different ways. As opposed to real nonsymmetric matrices, real symmetric matrices are always diagonalisable and have real eigenvalues (this is sometimes called the Spectral Theorem ) and thus each real symmetric matrix has its own inertia : the triple (p,n,z) where p is the number of positive eigenvalues, n is the number of negative eigenvalues and z is the number of zero eigenvalues. In this short article we consider an invertible nonsymmetric matrix {\sf B}\in{\mathbb{R}}^{m\times m} when we express it as {\sf B}={\sf T}{\sf W} with {\sf T},{\sf W}\in{\mathbb{R}}^{m\times m} being real symmetric. For clarity we employ the Parlett convention [ 5 , page 6] : symmetric characters refer to symmetric matrices. So {\sf T}={\sf T}^{T} and {\sf W}={\sf W}^{T} , but {\sf B}\neq{\sf B}^{T} are implied by this convention. Here we prove bounds relating the number of complex eigenvalues of a real invertible matrix {\sf B} and the inertia of real symmetric matrices {\sf T},{\sf W} which are such that {\sf B}={\sf T}{\sf W} . Note that because {\sf B} is invertible, then {\sf T}^{-1},{\sf W}^{-1} must exist. Further because \lambda>0\Leftrightarrow\frac{1}{\lambda}>0 , we have that the inertia of {\sf T}^{-1} is the same as the inertia of {\sf T} and the inertia of {\sf W}^{-1} is the same as the inertia of {\sf W} . We follow up with further comment on such factorisation and give several examples. Some of the results we present were known to Ostrowski [ 4 ] (see also [ 7 ] ). 1 Eigenvalues, Inertia and Bounds As soon as we learn about real polynomials, we encounter the possibility of complex roots. Applied to the Characteristic Polynomial we quickly discover that general (nonsymmetric) real matrices can have complex eigenvalues; real symmetric matrices on the other hand can only have real eigenvalues. In all cases the roots of a polynomial are continuously dependent on the coefficients, hence the eigenvalues of a matrix are continuously dependent on the matrix entries. Suppose {\sf B}={\sf T}{\sf W} , and note that {\sf T}^{-1},{\sf W}^{-1} must also be real symmetric. We then have that \lambda is an eigenvalue of {\sf B} \iff {{\sf B}-\lambda{\sf I}} is singular \iff {\sf T}{\sf W}-\lambda{\sf I} is singular \iff {\sf W}-\lambda{\sf T}^{-1} is singular. Now {\sf V}(\theta)=\theta{\sf W}+(1-\theta){\sf T}^{-1} is a real symmetric matrix for all \theta\in{\mathbb{R}} , hence {\sf V}(\theta) has real eigenvalues that depend continuously on \theta . Lemma 1.1 . If {\sf B}={\sf T}{\sf W} with {\sf T} and {\sf W} having different inertia then {\sf B} has at least one real negative eigenvalue. Proof. Since {\sf T} , thus {\sf T}^{-1}={\sf V}(0) and {\sf W}={\sf V}(1) have different inertia, at least one eigenvalue of {\sf V}(\theta) must change sign as \theta varies from 0 to 1 . Hence, by the Intemediate Value Theorem, there must exist \widehat{\theta}\in(0,1) such that {\sf V}(\widehat{\theta}\,) has a zero eigenvalue and therefore is singular. Thus {\sf W}+\frac{(1-\widehat{\theta}\,)}{\widehat{\theta}}\,{\sf T}^{-1}\quad{\mbox{and so also}}\quad{\sf T}{\sf W}+\frac{(1-\widehat{\theta}\,)}{\widehat{\theta}}\,{\sf I} are singular, hence \frac{(\widehat{\theta}-1)}{\widehat{\theta}}<0 is an eigenvalue of {\sf T}{\sf W}={\sf B} and the result follows. With a little more subtlety we can make a more precise quantitative statement. In all that follows we suppose that {\sf T} (and so {\sf T}^{-1} also) has inertia (p,n,0) and that {\sf W} (and so {\sf W}^{-1} also) has inertia (\widehat{p},\widehat{n},0) . Corollary 1.2 . {\sf B}={\sf T}{\sf W} has at least |p-\widehat{p}\,|=|n-\widehat{n}\,| real negative eigenvalues (counting multiplicities). Proof. The general principle of Avoidance of Crossing due to Lax [ 3 ] indicates why there are likely to be at least |p-\widehat{p}\,|=|n-\widehat{n}\,| values of \widehat{\theta}\in(0,1) for which {\sf V}({\widehat{\theta}}\,) is singular, though this is not quite sufficient to provide a general proof. A proof simply requires one to clarify mutiplicities if several eigenvalue trajectories of {\sf V}(\theta) change sign at the same value \widehat{\theta} . Note that for any such \widehat{\theta} , because of symmetry, the kernel of {\sf V}({\widehat{\theta}}\,) must possess an orthonormal basis {\cal{U}}=\{u_{1},\ldots,u_{r}\} where r is the nullity of {\sf V}({\widehat{\theta}}\,) . Thus \left({\widehat{\theta}}\;{\sf W}+(1-{\widehat{\theta}}\,)\,{\sf T}^{-1}\right)u_{i}=0, equivalently \left({\widehat{\theta}}\;{\sf T}{\sf W}+(1-{\widehat{\theta}}\,)\,{\sf I}\right)u_{i}=0 for i=1,\ldots,r . Dividing by {\widehat{\theta}} , it is seen that the vectors u_{1},\ldots,u_{r} are necessarily linearly independent right eigenvectors of {\sf B}={\sf T}{\sf W} corresponding to the eigenvalue \widehat{\lambda}=\frac{(\widehat{\theta}-1)}{\widehat{\theta}}<0 . Thus the kernel of {\sf B}-\widehat{\lambda}\,{\sf I} must be of dimension at least r when r eigenvalue trajectories of V(\theta) change sign at the same value \theta={\widehat{\theta}} . By applying similar argument to the above to the different parameterised real symmetric matrix {\sf U}(\phi)=\phi{\sf W}+(1-\phi)(-{\sf T}^{-1}) , noting that {\sf U}(1)={\sf W} and that {\sf U}(0)=-{\sf T}^{-1} which has inertia (n,p,0) we have Corollary 1.3 . The real matrix {\sf B}={\sf T}{\sf W} has at least |n-\widehat{p}\,|=|p-\widehat{n}\,| real and positive eigenvalues (counting multiplicities). Proof. Unless the inertia of {\sf W} and -{\sf T} are the same, there exist values \widehat{\phi}\in(0,1) such that {\sf U}(\widehat{\phi}\,) is singular. For any such \widehat{\phi} , as in the proof of the lemma above we have that \frac{1-{\widehat{\phi}}\,}{\widehat{\phi}} is a real, positive eigenvalue of {\sf B}={\sf T}{\sf W} . The same quantitative argument as in the corollary above gives the result. We now come to our main result. Because the complex numbers include the real numbers, we use ‘non-real’ to refer to complex numbers with non-vanishing imaginary parts. As above, the inertia of {\sf T} is (p,n,0) . Proposition 1.4 . Suppose {\sf B}\in{\mathbb{R}}^{m\times m} is invertible and {\sf B}={\sf T}{\sf W} with {\sf T}={\sf T}^{T}\in{\mathbb{R}}^{m\times m} , {\sf W}={\sf W}^{T}\in{\mathbb{R}}^{m\times m} . If {\sf B} has m-s non-real eigenvalues (thus \frac{1}{2}m-\frac{1}{2}s complex conjugate pairs) then \frac{1}{2}m-\frac{1}{2}s\leq p\leq\frac{1}{2}m+\frac{1}{2}s. Proof. We prove by contradiction: if p<\frac{1}{2}m-\frac{1}{2}s then n>\frac{1}{2}m+\frac{1}{2}s because p+n=m . Likewise if p>\frac{1}{2}m+\frac{1}{2}s then n<\frac{1}{2}m-\frac{1}{2}s . Thus, whatever the value of \widehat{p} we must have that |p-\widehat{p}\,|+|n-\widehat{p}\,|>\frac{1}{2}m+\frac{1}{2}s-\Bigl(\frac{1}{2}m-\frac{1}{2}s\Bigr)=s. (1) By the corollaries 1.2 and 1.3 above, there are at least |p-\widehat{p}\,| real negative eigenvalues and at least |n-\widehat{p}\,| real positive eigenvalues of {\sf B} , thus the inequality ( 1 ) implies that there must be greater than s real eigenvalues of {\sf B} . Because it has m eigenvalues in total (from the fundamental theorem of algebra), {\sf B} must have less than m-s non-real eigenvalues. This completes the proof by contradiction. A similar result applies to {\sf W} : Corollary 1.5 . If {\sf B}={\sf T}{\sf W} as above has m-s non-real eigenvalues then \frac{1}{2}m-\frac{1}{2}s\leq\widehat{p}\leq\frac{1}{2}m+\frac{1}{2}s. Proof. Note {\sf B}^{T}={\sf W}{\sf T} because {\sf T},{\sf W} are symmetric. The result follows from Proposition 1.4 by reversing the roles of {\sf T} and {\sf W} and noting that {\sf B}^{T} and {\sf B} are similar: {\sf