Maths WorkshopMaths problems with full solutions
Exercise 2028Topology › Topology of Mn(ℝ) › Rank, spectrum and normsM1Added 6 Nov 2026

Lyapunov's criterion for Hurwitz matrices

#Hurwitz#Lyapunov equation#stability#positive definite

Statement

A real matrix AA is *Hurwitz* if all its eigenvalues have negative real part. Show that AA is Hurwitz iff there exists a symmetric positive definite PP with ATP+PA<0A^TP+PA<0 (negative definite).

Hint

(⇐\Leftarrow) Evaluate the quadratic form on a complex eigenvector. (⇒\Rightarrow) Take P=∫0∞eATteAt dtP=\int_0^\infty e^{A^Tt}e^{At}\,dt.

Show solutionHide solution

Solution

(⇐\Leftarrow). Let Av=λvAv=\lambda v, v∈Cn∖{0}v\in\mathbb{C}^n\setminus\{0\}. Since AA is real, v∗AT=(Av)∗=λˉv∗v^*A^T=(Av)^*=\bar\lambda v^*. Therefore
v∗(ATP+PA)v=λˉ v∗Pv+λ v∗Pv=2Re⁡λ v∗Pv.v^*(A^TP+PA)v=\bar\lambda\,v^*Pv+\lambda\,v^*Pv=2\operatorname{Re}\lambda\ v^*Pv.
The left side is <0<0 (a negative definite real symmetric matrix is negative definite on Cn\mathbb{C}^n too), and v∗Pv>0v^*Pv>0. Hence Re⁡λ<0\operatorname{Re}\lambda<0.

(⇒\Rightarrow). Suppose AA is Hurwitz. Then there are C,c>0C,c>0 with ∥eAt∥≤Ce−ct\|e^{At}\|\le Ce^{-ct} for t≥0t\ge0 (Jordan form: entries of eAte^{At} are tjeλtt^je^{\lambda t} with Re⁡λ<0\operatorname{Re}\lambda<0). So the integral
P=∫0∞eATteAt dtP=\int_0^\infty e^{A^Tt}e^{At}\,dt
converges absolutely (integrand ≤C2e−2ct\le C^2e^{-2ct}). It is symmetric, and for x≠0x\ne0,
xTPx=∫0∞∣eAtx∣2dt>0,x^TPx=\int_0^\infty|e^{At}x|^2dt>0,
because eAte^{At} is invertible so eAtx≠0e^{At}x\ne0 for all tt and the integrand is continuous and positive. So P>0P>0. Moreover
ATP+PA=∫0∞ddt(eATteAt)dt=lim⁡T→∞eATTeAT−I=−I,A^TP+PA=\int_0^\infty\frac{d}{dt}\Big(e^{A^Tt}e^{At}\Big)dt=\lim_{T\to\infty}e^{A^TT}e^{AT}-I=-I,
which is negative definite.

Example. A=(−1100−1)A=\begin{pmatrix}-1&10\\0&-1\end{pmatrix} is Hurwitz but AT+AA^T+A is indefinite: P=IP=I does not work, yet a suitable P=(15551)P=\begin{pmatrix}1&5\\5&51\end{pmatrix}-like matrix does. The Lyapunov function V(x)=xTPxV(x)=x^TPx decreases along the flow x˙=Ax\dot x=Ax.

Remark. In topological terms, the set of Hurwitz matrices is open (eigenvalues are continuous), and it is the union over P>0P>0 of the open convex cones {A:ATP+PA<0}\{A:A^TP+PA<0\}; each cone is convex but the union is not (the set of Hurwitz matrices is *not* convex).