6.4 Linear Systems and Stability: Planar Systems and Lyapunov Functions
Linear systems $x'=Ax$ and the matrix exponential, phase portraits of planar systems classified by trace and determinant, Lyapunov stability and Lyapunov's direct method with full proofs, linearisation and the Hartman–Grobman theorem, and a first look at limit cycles.
Many systems in science are described by several quantities that change together: positions and velocities of coupled masses, concentrations in connected tanks, populations of predators and prey, currents in an electrical network. Their models are systems of first-order equations $x'=f(x)$ with $x\in\mathbb{R}^n$. When $f$ is linear, $f(x)=Ax$, there is a complete theory: the solution is $x(t)=e^{tA}x_0$, and the eigenvalues of $A$ decide everything about the long-term behaviour. When $f$ is nonlinear we usually cannot solve at all, but we can still answer the questions that matter in practice — where are the equilibria, and does a small disturbance die out or grow? Two tools answer them: linearisation, which replaces $f$ near an equilibrium by its Jacobian matrix, and Lyapunov functions, "energy-like" functions that decrease along solutions, which work even when linearisation is silent.
GATE MA tests this section with a steady pattern: compute $e^{tA}$ or one entry of it; solve a $2\times2$ or $3\times3$ system; classify the origin of $x'=Ax$ (node, saddle, spiral, centre; stable or not) from the trace and determinant, often with a parameter; find and classify the equilibria of a nonlinear system; decide, in an MSQ, which statements about stability, Lyapunov functions or linearisation are true; and occasionally locate a limit cycle in polar coordinates. Throughout, $\|\cdot\|$ is the Euclidean norm on $\mathbb{R}^n$ and the corresponding operator norm on matrices, $\tau=\operatorname{tr}A$, $\Delta=\det A$, and $D=\tau^2-4\Delta$ for $2\times2$ matrices.
Linear Systems and the Matrix Exponential
- The series for $e^M$ converges absolutely for every $M$, and $\|e^M\|\le e^{\|M\|}$.
- $t\mapsto e^{tA}$ is differentiable and $\dfrac{d}{dt}e^{tA}=Ae^{tA}=e^{tA}A$.
- If $AB=BA$, then $e^{A+B}=e^Ae^B$. In particular $e^{(t+s)A}=e^{tA}e^{sA}$, and $e^{tA}$ is invertible with inverse $e^{-tA}$.
- $e^{P^{-1}AP}=P^{-1}e^AP$; $e^{\operatorname{diag}(\lambda_1,\dots,\lambda_n)}=\operatorname{diag}(e^{\lambda_1},\dots,e^{\lambda_n})$; and $\det e^{tA}=e^{t\operatorname{tr}A}$.
- For every $x_0$, $x(t)=e^{(t-t_0)A}x_0$ is the unique solution of $x'=Ax$, $x(t_0)=x_0$; $e^{tA}$ is the fundamental matrix with $\Phi(0)=I$.
Proof. (1) The operator norm is submultiplicative, so $\|M^k/k!\|\le\|M\|^k/k!$, and $\sum\|M\|^k/k!=e^{\|M\|}$; absolute convergence in the finite-dimensional space of matrices implies convergence. (2) The termwise derivative of $\sum t^kA^k/k!$ is $\sum_{k\ge1}t^{k-1}A^k/(k-1)!=Ae^{tA}$. On $|t|\le T$ this derived series is dominated by $\sum\|A\|^kT^{k-1}/(k-1)!\lt\infty$, so it converges uniformly and termwise differentiation is justified (the interchange theorem of §4.2). The factor $A$ can be pulled out on either side. (3) Since $A$ and $B$ commute, the binomial theorem holds: $(A+B)^k=\sum_j\binom kjA^jB^{k-j}$. Multiplying the two absolutely convergent series (Cauchy product),
$$e^Ae^B=\sum_{k\ge0}\sum_{j=0}^{k}\frac{A^j}{j!}\frac{B^{k-j}}{(k-j)!}=\sum_{k\ge0}\frac{(A+B)^k}{k!}=e^{A+B}.$$The matrices $tA$ and $sA$ commute, and $e^{tA}e^{-tA}=e^{0}=I$. (4) $(P^{-1}AP)^k=P^{-1}A^kP$; sum the series. For a diagonal matrix the powers are diagonal. For the determinant, triangularise over $\mathbb{C}$ (§3.2): $A=PUP^{-1}$ with $U$ upper triangular with diagonal $\lambda_1,\dots,\lambda_n$. Every power $U^k$ is upper triangular with diagonal $\lambda_i^k$, so $e^{tU}$ is upper triangular with diagonal $e^{t\lambda_i}$, and $\det e^{tA}=\det e^{tU}=e^{t(\lambda_1+\cdots+\lambda_n)}=e^{t\operatorname{tr}A}$. (5) By (2), $x(t)=e^{(t-t_0)A}x_0$ is a solution, and $x(t_0)=x_0$. If $x$ is any solution, then by (2) and the product rule
$$\frac{d}{dt}\Big(e^{-(t-t_0)A}x(t)\Big)=-Ae^{-(t-t_0)A}x+e^{-(t-t_0)A}Ax=0,$$because $A$ commutes with $e^{-(t-t_0)A}$. So $e^{-(t-t_0)A}x(t)$ is constant, equal to its value $x_0$ at $t_0$, and $x(t)=e^{(t-t_0)A}x_0$. The columns of $e^{tA}$ are the solutions with $x(0)=e_j$, and $\det e^{tA}\ne0$ by (4). ∎
- Diagonalisable $A$. If $Av_i=\lambda_iv_i$ for a basis $v_1,\dots,v_n$ and $P=[v_1\cdots v_n]$, then $e^{tA}=P\operatorname{diag}(e^{\lambda_1t},\dots,e^{\lambda_nt})P^{-1}$, and every solution is $x(t)=\sum_ic_ie^{\lambda_it}v_i$.
- Jordan block. $e^{tJ_k(\lambda)}=e^{\lambda t}\sum_{j=0}^{k-1}\frac{t^j}{j!}N^j$ with $N=J_k(0)$: the $j$-th superdiagonal is $\frac{t^j}{j!}e^{\lambda t}$. For general $A=PJP^{-1}$ in Jordan form, $e^{tA}=Pe^{tJ}P^{-1}$.
- $2\times2$ formulas (no eigenvectors needed). If $A$ has eigenvalues $\lambda_1\ne\lambda_2$: $\ e^{tA}=\dfrac{e^{\lambda_1t}(A-\lambda_2I)-e^{\lambda_2t}(A-\lambda_1I)}{\lambda_1-\lambda_2}$. If $A$ has a double eigenvalue $\lambda$: $\ e^{tA}=e^{\lambda t}\big(I+t(A-\lambda I)\big)$. If $A$ is real with eigenvalues $\alpha\pm i\beta$, $\beta\gt0$: $$e^{tA}=e^{\alpha t}\left(\cos\beta t\;I+\frac{\sin\beta t}{\beta}\,(A-\alpha I)\right).$$
- Real solutions from complex eigenvalues. If $A$ is real and $A(a+ib)=(\alpha+i\beta)(a+ib)$, then $e^{\alpha t}(a\cos\beta t-b\sin\beta t)$ and $e^{\alpha t}(a\sin\beta t+b\cos\beta t)$ are real solutions.
Proof. (1) $A=PDP^{-1}$ and Theorem 6.42(4); expanding $e^{tA}x_0=Pe^{tD}c$ with $c=P^{-1}x_0$ gives $\sum c_ie^{\lambda_it}v_i$. (2) $J_k(\lambda)=\lambda I+N$, $\lambda I$ commutes with $N$, and $N^k=0$, so $e^{tJ}=e^{\lambda t}e^{tN}$ and the series for $e^{tN}$ stops. (3) By Cayley–Hamilton (§3.2), $(A-\lambda_1I)(A-\lambda_2I)=0$. Double eigenvalue: $(A-\lambda I)^2=0$, so $e^{t(A-\lambda I)}=I+t(A-\lambda I)$; multiply by $e^{\lambda tI}=e^{\lambda t}I$, which commutes. Complex pair: $p_A(x)=(x-\alpha)^2+\beta^2$, so $B=A-\alpha I$ satisfies $B^2=-\beta^2I$, hence $B^{2m}=(-1)^m\beta^{2m}I$ and $B^{2m+1}=(-1)^m\beta^{2m}B$. Splitting the series of $e^{tB}$ into even and odd terms gives $\cos\beta t\,I+\frac{\sin\beta t}{\beta}B$. Distinct eigenvalues: put $P_1=\frac{A-\lambda_2I}{\lambda_1-\lambda_2}$ and $P_2=\frac{A-\lambda_1I}{\lambda_2-\lambda_1}$. Then $P_1+P_2=I$, and $(A-\lambda_1I)P_1=0$, i.e. $AP_1=\lambda_1P_1$; similarly $AP_2=\lambda_2P_2$. Hence $A^kP_i=\lambda_i^kP_i$, $e^{tA}P_i=e^{\lambda_it}P_i$, and $e^{tA}=e^{tA}(P_1+P_2)=e^{\lambda_1t}P_1+e^{\lambda_2t}P_2$. (4) $x=e^{(\alpha+i\beta)t}(a+ib)$ solves $x'=Ax$; since $A$ is real, so do its real and imaginary parts, which are the two vectors stated. ∎
Proof. Write $x(t)=e^{tA}\big(x_0+\int_0^te^{-sA}b(s)\,ds\big)$. By the product rule and the fundamental theorem of calculus, $x'=Ax+e^{tA}e^{-tA}b(t)=Ax+b$, and $x(0)=x_0$. If $x,\tilde x$ are two solutions with the same initial value, $x-\tilde x$ solves $y'=Ay$, $y(0)=0$, so it is $0$ by Theorem 6.42(5). For constant $b$, $y=x-x^*$ satisfies $y'=Ax+b=A(x-x^*)=Ay$. ∎ (The Laplace transform of §6.2 gives the same results: $\mathcal{L}\{e^{tA}\}=(sI-A)^{-1}$.)
Phase Portraits and Stability of Linear Systems
- stable (in the sense of Lyapunov) if for every $\varepsilon\gt0$ there is $\delta\gt0$ such that every solution with $\|x(0)-x^*\|\lt\delta$ exists for all $t\ge0$ and satisfies $\|x(t)-x^*\|\lt\varepsilon$ for all $t\ge0$;
- asymptotically stable if it is stable and, in addition, there is $\delta_0\gt0$ such that $\|x(0)-x^*\|\lt\delta_0$ implies $x(t)\to x^*$ as $t\to\infty$;
- unstable if it is not stable.
- asymptotically stable if and only if $\operatorname{Re}\lambda_i\lt0$ for every $i$; then there are $C\ge1$ and $\mu\gt0$ with $\|e^{tA}\|\le Ce^{-\mu t}$ for $t\ge0$, so every solution tends to $0$ exponentially;
- stable if and only if $\operatorname{Re}\lambda_i\le0$ for every $i$ and every eigenvalue with $\operatorname{Re}\lambda=0$ is semisimple (all its Jordan blocks are $1\times1$; equivalently its algebraic and geometric multiplicities agree);
- unstable otherwise: if some $\operatorname{Re}\lambda\gt0$, or some eigenvalue with $\operatorname{Re}\lambda=0$ has a Jordan block of size $\ge2$.
Proof. Write $A=PJP^{-1}$ in Jordan form over $\mathbb{C}$ (§3.2). By Theorem 6.43(2), every entry of $e^{tJ}$ is $0$ or of the form $\frac{t^j}{j!}e^{\lambda t}$, where $\lambda$ is the eigenvalue of its block and $j$ is less than the block size. Since $\|e^{tA}\|\le\|P\|\,\|e^{tJ}\|\,\|P^{-1}\|$ and $|t^je^{\lambda t}|=t^je^{(\operatorname{Re}\lambda)t}$:
Sufficiency in (1). Choose $\mu\gt0$ with $\operatorname{Re}\lambda_i\lt-2\mu$ for all $i$. Then $t^je^{(\operatorname{Re}\lambda)t}\le t^je^{-2\mu t}\le C_je^{-\mu t}$, because $t^je^{-\mu t}$ is bounded on $[0,\infty)$. So $\|e^{tA}\|\le Ce^{-\mu t}$, which gives stability (take $\delta=\varepsilon/C$) and convergence to $0$. Sufficiency in (2). Entries with $\operatorname{Re}\lambda\lt0$ are bounded as before; an eigenvalue with $\operatorname{Re}\lambda=0$ and only $1\times1$ blocks contributes entries $e^{\lambda t}$ of modulus $1$. So $\|e^{tA}\|\le C$ for $t\ge0$ and $\delta=\varepsilon/C$ works.
Necessity (and (3)). If $Av=\lambda v$ with $\operatorname{Re}\lambda\gt0$, $v\ne0$, then $x(t)=\operatorname{Re}\big(e^{\lambda t}\varepsilon v\big)$ is a real solution starting within $\varepsilon\|v\|$ of $0$ whose size grows like $\varepsilon e^{(\operatorname{Re}\lambda)t}$ (unboundedly, along a sequence of times): unstable. If $\operatorname{Re}\lambda=0$ and there is a Jordan chain $(A-\lambda I)w=v$, $(A-\lambda I)v=0$, then $x(t)=e^{\lambda t}(w+tv)$ solves $x'=Ax$ (differentiate), and the real or imaginary part of $\varepsilon x(t)$ starts near $0$ and grows like $\varepsilon t$: unstable. Finally, if some $\operatorname{Re}\lambda=0$, the solution $\operatorname{Re}(e^{\lambda t}\varepsilon v)$ does not tend to $0$, so $0$ is not asymptotically stable. ∎
- saddle if the eigenvalues are real with opposite signs;
- node if they are real with the same sign — a star (proper node) if $A=\lambda I$, a degenerate (improper) node if $\lambda_1=\lambda_2$ but $A\ne\lambda I$;
- spiral (focus) if they are $\alpha\pm i\beta$ with $\alpha\ne0\ne\beta$;
- centre if they are $\pm i\beta$ with $\beta\ne0$.
Let $A=\begin{pmatrix}a&b\\c&d\end{pmatrix}$ be real, $\tau=a+d$, $\Delta=ad-bc\ne0$, $D=\tau^2-4\Delta$. The eigenvalues are $\frac12(\tau\pm\sqrt D)$, and:
| condition | eigenvalues | type of $0$ | stability |
|---|---|---|---|
| $\Delta\lt0$ | real, opposite signs | saddle | unstable |
| $\Delta\gt0$, $D\gt0$ | real, distinct, sign of $\tau$ | node | asympt. stable iff $\tau\lt0$ |
| $\Delta\gt0$, $D=0$ | double, $\tau/2$ | star or degenerate node | asympt. stable iff $\tau\lt0$ |
| $\Delta\gt0$, $D\lt0$, $\tau\ne0$ | $\frac\tau2\pm i\frac{\sqrt{-D}}{2}$ | spiral | asympt. stable iff $\tau\lt0$ |
| $\Delta\gt0$, $\tau=0$ | $\pm i\sqrt\Delta$ | centre | stable, not asympt. |
Consequently $0$ is asymptotically stable iff $\tau\lt0$ and $\Delta\gt0$. Moreover: (i) spirals and centres turn anticlockwise if $c\gt0$ and clockwise if $c\lt0$; (ii) for a centre, the orbits are the ellipses $cx^2-2axy-by^2=\text{const}$, and every solution is periodic with period $2\pi/\sqrt\Delta$; (iii) at a node with $|\lambda_1|\lt|\lambda_2|$, every orbit not on the eigenline of $\lambda_2$ enters (or leaves) the origin tangent to the eigenvector of the slow eigenvalue $\lambda_1$.
Proof. $p_A(\lambda)=\lambda^2-\tau\lambda+\Delta$, so $\lambda_1\lambda_2=\Delta$ and $\lambda_1+\lambda_2=\tau$. If $\Delta\lt0$, then $D\gt0$ and the real roots have product $\lt0$. If $\Delta\gt0$ and $D\ge0$, the real roots have product $\gt0$, hence a common sign, the sign of their sum $\tau$. If $D\lt0$ the roots are $\frac\tau2\pm\frac i2\sqrt{-D}$. The stability column is Theorem 6.45 (for a centre the eigenvalues $\pm i\sqrt\Delta$ are distinct, hence semisimple). (i) On the positive $x$-axis, $y'=cx$ has the sign of $c$. (ii) With $d=-a$ and $Q=cx^2-2axy-by^2$, a direct computation gives $\frac{d}{dt}Q(x,y)=2cxx'-2a(x'y+xy')-2byy'$ with $x'=ax+by$, $y'=cx-ay$; expanding, the terms $\pm2acx^2$, $\pm2bcxy$, $\pm2a^2xy$ and $\pm2aby^2$ cancel in pairs, so $\frac{d}{dt}Q=0$, and $Q$ is definite because $\det\begin{pmatrix}c&-a\\-a&-b\end{pmatrix}=-bc-a^2=\Delta\gt0$; its level curves are ellipses. The eigenvalues $\pm i\sqrt\Delta$ give period $2\pi/\sqrt\Delta$. (iii) $x(t)=c_1e^{\lambda_1t}v_1+c_2e^{\lambda_2t}v_2$; at a stable node, as $t\to\infty$, the term $e^{\lambda_1t}$ decays more slowly and dominates when $c_1\ne0$, so $x(t)/\|x(t)\|\to\pm v_1/\|v_1\|$ (for an unstable node let $t\to-\infty$). ∎
- Compute $\tau$, $\Delta$ and, if $\Delta\gt0$, $D=\tau^2-4\Delta$. Read the type from the table; check $\Delta=0$ separately (then a whole line of equilibria, $N(A)$).
- Real eigenvalues: find the eigenvectors, draw the eigenlines, mark the motion on them (out for $\lambda\gt0$, in for $\lambda\lt0$); other orbits are tangent to the slow direction near $0$ and parallel to the fast direction far away.
- Complex eigenvalues: the sign of $\tau$ gives in/out, the sign of $c$ gives the sense of rotation; for a centre draw ellipses.
- With a parameter, sketch the path of $(\tau,\Delta)$ in the trace–determinant plane against the parabola $\Delta=\tau^2/4$ (Interactive 1).
Lyapunov Functions and Lyapunov's Theorems
For $x'=Ax$ the eigenvalues settle stability. For nonlinear systems Lyapunov's idea (1892) is to imitate the physicist's energy: if some positive quantity can only decrease along every solution, solutions near a minimum of it are trapped. The point is that the rate of change of $V(x(t))$ can be computed without solving the system.
- If $\dot V\le0$ on $U$, then $0$ is stable.
- If $\dot V$ is negative definite on $U$, then $0$ is asymptotically stable.
Proof. (1) Let $\varepsilon\gt0$ with the closed ball $\bar B_\varepsilon\subset U$. The sphere $S_\varepsilon=\{\|x\|=\varepsilon\}$ is compact and $V\gt0$ on it, so $m=\min_{S_\varepsilon}V\gt0$. By continuity and $V(0)=0$ there is $\delta\in(0,\varepsilon)$ with $V(x)\lt m$ for $\|x\|\lt\delta$. Let $\|x(0)\|\lt\delta$. As long as the solution stays in $\bar B_\varepsilon$, $V(x(t))$ is non-increasing, so $V(x(t))\le V(x(0))\lt m$; therefore the solution can never reach $S_\varepsilon$, where $V\ge m$. It stays in the compact set $\bar B_\varepsilon$, hence exists for all $t\ge0$ (a solution can only stop existing by leaving every compact set), and $\|x(t)\|\lt\varepsilon$.
(2) Take $\delta_0=\delta$ from (1) for some fixed $\varepsilon$. Since $V(x(t))$ decreases and is $\ge0$, it has a limit $L\ge0$. Suppose $L\gt0$. By continuity of $V$ there is $\rho\gt0$ with $V\lt L$ on $B_\rho$, so the solution, on which $V\ge L$, stays in the compact annulus $K=\{\rho\le\|x\|\le\varepsilon\}$. On $K$ the continuous negative function $\dot V$ has a maximum $-\gamma\lt0$, so $V(x(t))\le V(x(0))-\gamma t\to-\infty$, contradicting $V\ge0$. So $L=0$. Finally $x(t)\to0$: otherwise $\|x(t_k)\|\ge\rho'\gt0$ along a sequence $t_k\to\infty$, and $V(x(t_k))\ge\min\{V(x):\rho'\le\|x\|\le\varepsilon\}\gt0$, contradicting $L=0$. ∎
Proof. Fix $\varepsilon\gt0$ with $\bar B_\varepsilon\subset U$; we show that solutions starting arbitrarily close to $0$ leave $\bar B_\varepsilon$. Given $\delta\gt0$, choose $x_0$ with $\|x_0\|\lt\delta$ and $V(x_0)=v_0\gt0$. Suppose the solution stays in $\bar B_\varepsilon$ for all $t\ge0$. Since $\dot V\ge0$, $V(x(t))\ge v_0$. The set $K=\{x\in\bar B_\varepsilon:V(x)\ge v_0\}$ is compact and does not contain $0$, so $\dot V\ge\gamma\gt0$ on $K$, and $V(x(t))\ge v_0+\gamma t\to\infty$ — impossible, since $V$ is bounded on $\bar B_\varepsilon$. ∎
Proof (sketch). A bounded forward orbit has a nonempty compact $\omega$-limit set $\omega$ (all limits of $x(t_k)$ with $t_k\to\infty$), which is invariant and is approached by the orbit — standard facts proved with Bolzano–Weierstrass and continuous dependence on initial data. $V(x(t))$ decreases to a limit $L$, so $V\equiv L$ on $\omega$. A solution starting in $\omega$ stays in $\omega$, so $V$ is constant along it and $\dot V=0$ there: $\omega\subseteq M$. For the corollary, Theorem 6.47(1) gives stability, a small sublevel set of $V$ around $0$ is a compact positively invariant $\Omega$, and $M=\{0\}$ by hypothesis. ∎
Proof. By Theorem 6.45(1), $\|e^{sA}\|\le Ce^{-\mu s}$, so the integrand is bounded by $C^2\|Q\|e^{-2\mu s}$ and the integral converges. $P$ is symmetric, and $x^{\mathsf T}Px=\int_0^\infty(e^{sA}x)^{\mathsf T}Q(e^{sA}x)\,ds\gt0$ for $x\ne0$, since $e^{sA}x\ne0$. Next, $\frac{d}{ds}\big(e^{sA^{\mathsf T}}Qe^{sA}\big)=A^{\mathsf T}e^{sA^{\mathsf T}}Qe^{sA}+e^{sA^{\mathsf T}}Qe^{sA}A$, and integrating over $[0,\infty)$ gives $A^{\mathsf T}P+PA=\big[e^{sA^{\mathsf T}}Qe^{sA}\big]_0^\infty=-Q$. Uniqueness: if $A^{\mathsf T}X+XA=0$, then $\frac{d}{ds}\big(e^{sA^{\mathsf T}}Xe^{sA}\big)=0$, so $X=e^{sA^{\mathsf T}}Xe^{sA}\to0$, i.e. $X=0$. Along solutions $\dot V=x'^{\mathsf T}Px+x^{\mathsf T}Px'=x^{\mathsf T}(A^{\mathsf T}P+PA)x=-x^{\mathsf T}Qx$. Conversely, $V=x^{\mathsf T}Px$ is then a strict Lyapunov function, $0$ is asymptotically stable by Theorem 6.47, and Theorem 6.45(1) gives the eigenvalue condition. ∎
- Energy first. For $x''+h(x')+g(x)=0$ (as $x'=y$, $y'=-g(x)-h(y)$) try $V=\frac12y^2+\int_0^xg$; then $\dot V=-y\,h(y)$.
- Quadratic or even-power trials. Try $V=ax^2+by^2$, or $ax^{2p}+by^{2q}$, and choose $a,b,p,q\gt0$ so that the cross terms in $\dot V$ cancel. The cubic and quintic terms in the system tell you which powers to use.
- Semidefinite $\dot V$? Use LaSalle: find where $\dot V=0$ and check which solutions can stay there.
- Instability. Look for $V$ with $\dot V\gt0$ and $V\gt0$ somewhere near $0$ (Theorem 6.48); $x^2-y^2$ or $xy$ are typical near saddles.
- Failing to find a Lyapunov function proves nothing.
Linearisation, Hartman–Grobman and Limit Cycles
- If every eigenvalue of $A$ has negative real part, $x^*$ is asymptotically stable (indeed exponentially: $\|x(t)-x^*\|\le Ce^{-\mu t}\|x(0)-x^*\|$ near $x^*$).
- If some eigenvalue of $A$ has positive real part, $x^*$ is unstable.
- If $A$ has an eigenvalue with zero real part and none with positive real part, the linearisation decides nothing: the nonlinear terms do.
Proof. Put $x^*=0$ and write $f(x)=Ax+g(x)$; since $f$ is $C^1$ and $g(0)=0$, $Dg(0)=0$, we have $\|g(x)\|/\|x\|\to0$ as $x\to0$. (1) Let $P$ solve $A^{\mathsf T}P+PA=-I$ (Theorem 6.50) and $V=x^{\mathsf T}Px$. Along solutions of the nonlinear system,
$$\dot V=x^{\mathsf T}(A^{\mathsf T}P+PA)x+2x^{\mathsf T}Pg(x)\le-\|x\|^2+2\|P\|\,\|x\|\,\|g(x)\| .$$Choose $r\gt0$ with $\|g(x)\|\le\frac{\|x\|}{4\|P\|}$ for $\|x\|\le r$. There $\dot V\le-\frac12\|x\|^2$, negative definite, and Theorem 6.47(2) gives asymptotic stability. With $\lambda_{\min},\lambda_{\max}$ the extreme eigenvalues of $P$, $\dot V\le-\frac{1}{2\lambda_{\max}}V$, so $V(x(t))\le e^{-t/(2\lambda_{\max})}V(x(0))$, which is the exponential estimate. (2) Sketch: after a linear change of coordinates $x=(u,s)$ separating the eigenvalues with positive real part from the others, one builds quadratic forms $P_u,P_s$ (from Lyapunov equations for the two blocks, after shifting the second block slightly) such that $V=u^{\mathsf T}P_uu-s^{\mathsf T}P_ss$ has $\dot V\ge c\|x\|^2$ near $0$; $V\gt0$ on the $u$-axis arbitrarily close to $0$, and Chetaev's form of Theorem 6.48 gives instability. (3) Example 4: (a) and (b) have the same centre as linearisation, but one is asymptotically stable and the other unstable; Exercise IV.3 gives a third. ∎
Two remarks. A homeomorphism cannot tell a stable node from a stable spiral (both are "sinks"), which is why the $C^1$ refinement is needed to transfer the finer type. And Hartman–Grobman does not say how large $U$ is: the basin of attraction of a stable equilibrium is a global question, answered by Lyapunov functions and LaSalle (as in Example 5), not by linearisation.
The closed orbits around a linear centre are not limit cycles: they come in a continuous family. A limit cycle is a genuinely nonlinear phenomenon — a self-sustained oscillation of a fixed amplitude, like the beating of a heart, the squeal of a brake or the oscillation of a valve circuit described by van der Pol's equation $x''-\mu(1-x^2)x'+x=0$.
Proof. Suppose $\Gamma\subset D$ is a periodic orbit of period $T$; it is a simple closed curve (orbits of an autonomous system cannot cross themselves), and since $D$ is simply connected, the region $R$ enclosed by $\Gamma$ lies in $D$. Orient $\Gamma$ positively (replacing $t$ by $-t$ if necessary, which changes nothing below). By Green's theorem (§2.2),
$$\iint_R\operatorname{div}(\varphi f)\,dA=\oint_\Gamma\varphi f_1\,dy-\varphi f_2\,dx=\int_0^T\varphi\,\big(f_1\,y'-f_2\,x'\big)\,dt=\int_0^T\varphi\,(f_1f_2-f_2f_1)\,dt=0,$$using $x'=f_1$, $y'=f_2$ along $\Gamma$. But the integrand on the left is continuous and of one strict sign on $R$, which has positive area, so the left side is not $0$. ∎
The usual way to apply the theorem is a trapping annulus: a ring whose boundary the flow crosses inwards on both edges and which contains no equilibrium. The proof uses the Jordan curve theorem and the fact that a planar orbit cannot cross itself; it fails in dimension three (chaos is possible there — the Lorenz equations), which is why limit-cycle theory is special to the plane.
- $e^{A+B}\ne e^Ae^B$ in general. For $A=\begin{pmatrix}0&1\\0&0\end{pmatrix}$, $B=\begin{pmatrix}0&0\\1&0\end{pmatrix}$: $e^Ae^B=\begin{pmatrix}2&1\\1&1\end{pmatrix}$ but $e^{A+B}=\begin{pmatrix}\cosh1&\sinh1\\\sinh1&\cosh1\end{pmatrix}$. Commutativity is essential.
- $e^{tA}$ is not the entrywise exponential. $\begin{pmatrix}e^{ta_{ij}}\end{pmatrix}$ equals $e^{tA}$ only for diagonal $A$.
- One invariant is never enough in the plane. $\operatorname{tr}A\lt0$ alone fails ($\operatorname{diag}(1,-3)$ is a saddle); $\det A\gt0$ alone fails ($I$ is an unstable star). Both together are needed. In dimension $3$, $\operatorname{tr}A\lt0$ and $\det A\lt0$ are necessary but not sufficient: $\operatorname{diag}(1,1,-3)$.
- $\operatorname{Re}\lambda\le0$ does not give stability. $x'=y$, $y'=0$ (eigenvalues $0,0$, one Jordan block) has the unbounded solutions $(x_0+y_0t,\,y_0)$. Semisimplicity of the imaginary-axis eigenvalues is needed (Theorem 6.45(2)).
- Stable $\ne$ asymptotically stable (centres), and attracting $\ne$ stable: there are planar systems (Vinograd's example) in which every solution tends to the equilibrium but some first make large excursions from arbitrarily close starting points. That is why Definition 6.22 demands both properties.
- A linear centre decides nothing for the nonlinear system (Example 6(a)), and neither does a zero eigenvalue: $x'=-x^3$, $y'=-y$ is asymptotically stable and $x'=x^2$, $y'=-y$ is unstable, with the same linearisation.
- A Lyapunov function must be positive definite. For $x'=-x$, $y'=y$, the function $V=x^2$ has $\dot V=-2x^2\le0$, yet the origin is unstable: $V$ vanishes on the $y$-axis, which is exactly where solutions escape.
- $\dot V\le0$ gives only stability; for asymptotic stability use negative definiteness or LaSalle (Example 5). And not finding a Lyapunov function proves nothing.
- Bendixson–Dulac is a one-way test. If $\operatorname{div}f$ changes sign, nothing follows (Example 7(b) needed a weight). If $\operatorname{div}f\equiv0$ (e.g. a centre $x'=-y$, $y'=x$) there may be infinitely many periodic orbits.
- Hartman–Grobman is local and needs hyperbolicity. It says nothing about the size of the basin of attraction, nothing at a centre, and on its own it does not distinguish nodes from spirals.
- CSIR Unit 3 · §1.3 Systems of First-Order Equations — reduction of $n$-th order equations to systems and existence for systems, fundamental matrices and Liouville's formula for variable coefficients, the matrix exponential with generalised eigenvectors worked out in dimension three, Duhamel's formula, the trace–determinant classification, and a proof of linearised stability through Duhamel's formula and Gronwall's inequality (an alternative to the Lyapunov-function proof given here).
Two Labs
Exercises
- 1 Compute $e^{tA}$ for $A=\begin{pmatrix}0&1\\-4&0\end{pmatrix}$.
- 2 Compute $e^{tA}$ for $A=\begin{pmatrix}2&1\\0&2\end{pmatrix}$ directly from the series, and check with Theorem 6.43(3).
- 3 Compute $e^{tA}$ for $A=\begin{pmatrix}1&2\\2&1\end{pmatrix}$ using the distinct-eigenvalue formula.
- 4 Solve $x'=Ax$, $x(0)=(1,1,1)^{\mathsf T}$, for $A=\begin{pmatrix}1&1&0\\0&1&0\\0&0&2\end{pmatrix}$.
- 5 Find $\det e^{A}$ for $A=\begin{pmatrix}1&2\\3&4\end{pmatrix}$ without computing $e^A$.
- 6 Verify the first pitfall: compute $e^A$, $e^B$, $e^Ae^B$ and $e^{A+B}$ for $A=\begin{pmatrix}0&1\\0&0\end{pmatrix}$, $B=\begin{pmatrix}0&0\\1&0\end{pmatrix}$.
- 7 Solve $x'=\begin{pmatrix}0&1\\-1&0\end{pmatrix}x+\begin{pmatrix}0\\1\end{pmatrix}$, $x(0)=0$.
- 8 (Two tanks.) Two well-stirred tanks hold $100$ L of water each. Water is pumped from tank 1 to tank 2 and from tank 2 to tank 1 at $10$ L/min each way. Initially tank 1 contains $10$ kg of dissolved salt and tank 2 none. Find the salt amounts $x_1(t),x_2(t)$ and the time when tank 2 holds $4$ kg.
- 1 Classify the origin for $A=\begin{pmatrix}1&3\\3&1\end{pmatrix}$ and give the eigenlines.
- 2 Classify the origin for $A=\begin{pmatrix}-2&1\\-1&-2\end{pmatrix}$ and give the sense of rotation.
- 3 Show that $A=\begin{pmatrix}0&2\\-8&0\end{pmatrix}$ gives a centre, find a conserved quadratic and the period.
- 4 Classify the origin for $A=\begin{pmatrix}-3&1\\-1&-1\end{pmatrix}$.
- 5 (Damping.) For which $\alpha$ is the origin of the system form of $x''+\alpha x'+4x=0$ asymptotically stable? For which is it a spiral, a degenerate node, a node?
- 6 Decide stability for $A=\begin{pmatrix}-1&1&0\\0&-1&1\\0&0&-1\end{pmatrix}$, $B=\begin{pmatrix}0&1&0\\0&0&0\\0&0&-1\end{pmatrix}$ and $C=\operatorname{diag}(0,0,-1)$.
- 7 Prove that for real $2\times2$ $A$ the origin is asymptotically stable iff $\operatorname{tr}A\lt0$ and $\det A\gt0$. Show by an example that the analogous statement with $\operatorname{tr}A\lt0$, $\det A\lt0$ fails for $3\times3$ matrices.
- 8 Show that if $A$ is real skew-symmetric, every solution of $x'=Ax$ has constant norm. Deduce that $0$ is stable but not asymptotically stable (for $A\ne0$ and also for $A=0$).
- 1 Use $V=x^2+y^2$ to show that the origin of $x'=-x+y$, $y'=-x-y^3$ is asymptotically stable.
- 2 For $x'=-x^3+2y^3$, $y'=-2xy^2$, find $V=ax^2+by^2$ with $\dot V\le0$, and use LaSalle to show that the origin is asymptotically stable. What does linearisation say?
- 3 Find $c\gt0$ such that $V=x^2+cy^2$ has positive definite $\dot V$ for $x'=2xy+x^3$, $y'=-x^2+y^5$, and conclude.
- 4 Show that the origin of $x''+x^3=0$ is stable but not asymptotically stable.
- 5 Solve the Lyapunov equation $A^{\mathsf T}P+PA=-I$ for $A=\begin{pmatrix}-1&1\\0&-2\end{pmatrix}$ and check that $P$ is positive definite.
- 6 Let $V$ be a Lyapunov function on $U$ and $c\gt0$ such that the component $\Omega_c$ of $\{V\le c\}$ containing $0$ is compact and contained in $U$. Prove that $\Omega_c$ is positively invariant.
- 7 Show that the origin of $x''+x'+x^3=0$ is asymptotically stable, although the linearisation has a zero eigenvalue.
- 8 Prove or disprove: if $V$ is positive definite and $\dot V$ is negative semidefinite, the origin is asymptotically stable.
- 1 (Competition.) Find and classify all equilibria of $x'=x(4-x-y)$, $y'=y(6-x-3y)$, $x,y\ge0$. Which outcome do you expect for two competing species?
- 2 Find and classify the equilibria of $x'=y$, $y'=-x+x^3-y$.
- 3 Show that for $x'=-y-x(x^2+y^2)^2$, $y'=x-y(x^2+y^2)^2$ the linearisation is a centre but the origin is asymptotically stable.
- 4 Find and classify the equilibria of $x'=x^2-y^2$, $y'=x-1$. Are they hyperbolic?
- 5 Show that $x'=y$, $y'=-x-y+x^2$ has no periodic orbits.
- 6 Use the Dulac function $\varphi=\frac{1}{xy}$ to show that $x'=x(2-x-y)$, $y'=y(4x-x^2-3)$ has no periodic orbits in the open first quadrant.
- 7 Find all periodic orbits of $x'=-y+x(1-r^2)(4-r^2)$, $y'=x+y(1-r^2)(4-r^2)$, $r^2=x^2+y^2$, and decide their stability.
- 8 Show that $x'=1+x^2+y^2$, $y'=xy$ has no periodic orbit, in two ways.
- 9 For van der Pol's equation $x''-\mu(1-x^2)x'+x=0$ with $\mu\gt0$, classify the origin. Why must any periodic orbit enclose the origin?
E1. GATE MA · MCQ For $A=\begin{pmatrix}2&1\\-1&0\end{pmatrix}$, the $(1,2)$ entry of $e^{tA}$ is (A) $te^{t}$; (B) $e^{t}$; (C) $te^{2t}$; (D) $(1+t)e^{t}$.
E2. GATE MA · MCQ For the predator–prey system $x'=x(2-y)$, $y'=y(x-3)$ in the open first quadrant, the equilibrium $(3,2)$ is (A) an asymptotically stable spiral; (B) an unstable spiral; (C) a centre; (D) a saddle.
E3. GATE MA · MSQ Consider $x'=-x^3+y$, $y'=-x-y^3$. Which of the following is/are true? (One or more options may be correct.) (A) The linearisation at the origin is a centre. (B) The origin is asymptotically stable. (C) The stability of the origin can be decided from the eigenvalues of the Jacobian at the origin. (D) Along every solution, $x(t)^2+y(t)^2$ is non-increasing.
E4. GATE MA · MSQ Let $A$ be a real $2\times2$ matrix and consider $x'=Ax$. Which of the following is/are true? (One or more options may be correct.) (A) If $\det A\lt0$, the origin is unstable. (B) If $\operatorname{tr}A\lt0$, the origin is asymptotically stable. (C) If $\operatorname{tr}A=0$ and $\det A\gt0$, every nonzero solution is periodic with period $2\pi/\sqrt{\det A}$. (D) If $A^2=0$ and $A\ne0$, the origin is stable.
E5. GATE MA · NAT The system $x'=2x-y-x(x^2+y^2)$, $y'=x+2y-y(x^2+y^2)$ has exactly one periodic orbit, a circle centred at the origin. Its radius is ______ (correct to two decimal places).
E6. GATE MA · NAT Let $x(t)=(x_1(t),x_2(t))$ solve $x'=\begin{pmatrix}0&1\\-2&-3\end{pmatrix}x$, $x(0)=(1,0)$. Then $x_1(1)$ equals ______ (correct to two decimal places).
Suggested plan — 3 lectures × 60 min.
- The matrix exponential. Definition 6.20, Theorem 6.42 with the uniqueness proof (the derivative of $e^{-tA}x(t)$ is the key two lines), Theorem 6.43 with the Cayley–Hamilton derivation of the three $2\times2$ formulas, Theorem 6.44. Examples 1 and 2 on the board; GATE corner E1 and E6 as a quick check.
- Stability and phase portraits. Definitions 6.21–6.23, Theorem 6.45 (prove sufficiency fully; the counterexample $x'=y,y'=0$ for necessity), Theorem 6.46 with the table. Project Interactive 1: walk the red dot through every region and across the parabola. Example 3 as class work; E4 as a discussion.
- Lyapunov and linearisation. Definition 6.24, Theorems 6.47–6.48 with proofs (draw the ball, the sphere where $V\ge m$, and the smaller ball where $V\lt m$), LaSalle and Example 5, Theorem 6.50 and the Lyapunov-function proof of Theorem 6.51(1). Example 4 with Interactive 2 (systems 1–3). Close with Hartman–Grobman, Bendixson–Dulac and Example 7, then system 5 of Interactive 2.
Discussion questions. (1) Why does Theorem 6.45 need the semisimplicity condition only for eigenvalues on the imaginary axis, and never for eigenvalues with negative real part? (2) A Lyapunov function is like an energy, but the damped pendulum's energy has $\dot E=0$ on a whole line. What extra idea rescues asymptotic stability, and what goes wrong if the friction acts only on part of the line? (3) Why can limit cycles exist only for nonlinear systems, and why does the Poincaré–Bendixson theorem fail in three dimensions?
Misconceptions to watch for. $e^{A+B}=e^Ae^B$ without commuting; entrywise exponentials; "trace negative ⇒ stable"; "$\operatorname{Re}\lambda\le0$ ⇒ stable"; reading a nonlinear centre off a linear one; treating $\dot V\le0$ as asymptotic stability; accepting a semidefinite $V$ as a Lyapunov function; using Bendixson when the divergence changes sign; believing the basin of attraction is "small" or "everything" because of linearisation.
Suggested homework. I(1, 3, 4, 7, 8), II(2, 3, 5, 6, 7), III(2, 3, 5, 7), IV(1, 3, 6, 7, 9); GATE corner E2, E4, E5.
Research question. Can enriching an ecosystem destabilise it? If the food supply of a prey species is increased, does a stable predator–prey balance turn into boom-and-bust cycles — and at which level of enrichment?
Why it matters. Fertiliser run-off into lakes, fish stocking and nutrient-rich aquaculture ponds all "enrich" ecosystems. Rosenzweig (1971) predicted from a simple model that enrichment can produce violent oscillations and near-extinctions — the paradox of enrichment — and laboratory plankton experiments have since been used to test it. The mathematics is exactly this section: an equilibrium loses stability when the trace of its Jacobian changes sign, and a limit cycle is born (a Hopf bifurcation).
What to do.
- Take the dimensionless model $x'=x\left(1-\frac xK\right)-\frac{xy}{1+x}$, $y'=\frac{xy}{1+x}-\frac12y$ (prey $x$ with carrying capacity $K$, predators $y$ with a saturating feeding rate). Show that the coexistence equilibrium is $x^*=1$, $y^*=2\left(1-\frac1K\right)$ for $K\gt1$.
- Compute the Jacobian there and show that $\Delta\gt0$ and $\tau=\frac12-\frac{3}{2K}$. Conclude, using Theorems 6.46 and 6.51, that the equilibrium is asymptotically stable for $1\lt K\lt3$ and unstable for $K\gt3$.
- For $K=2.5,\ 3.5,\ 4,\ 6$ integrate numerically from $(1.2,1)$ for a long time and plot the orbits in the $(x,y)$ plane. For $K\gt3$ record the minimum and maximum prey density on the limit cycle and its period.
- Interpret: what happens to the minimum prey density as $K$ grows, and why is that dangerous for a real population that is subject to random fluctuations?
Tools: Python with scipy.integrate.solve_ivp (or any ODE solver); Interactive 1 to see the trace crossing zero.
Hand in: the equilibrium and stability computation; the four phase portraits; a table of cycle amplitude and period against $K$; a short paragraph on the ecological meaning.
Going further. Read about the Hopf bifurcation theorem, which guarantees the birth of a small limit cycle when a pair of complex eigenvalues crosses the imaginary axis with nonzero speed, and about the debate on whether the paradox of enrichment is seen in nature.
🧪 Self-check quiz
Q1. A real $2\times2$ matrix $A$ has $\operatorname{tr}A=-3$ and $\det A=2$. The origin of $x'=Ax$ is
Q2. If $\operatorname{tr}A=2$, then $\det e^{A}$ equals
Q3. For $x'=-y$, $y'=x$ and $V=x^2+y^2$, which conclusion is correct?
Q4. The Hartman–Grobman theorem applies at an equilibrium $x^*$ when
Q5. A planar $C^1$ system satisfies $\operatorname{div}f\lt0$ at every point of $\mathbb{R}^2$. Then