1.3 Systems of First-Order Equations
Vector equations $x'=f(t,x)$: reduction of order-$n$ equations, existence and uniqueness, fundamental matrices and Liouville's formula, $x'=Ax$ through eigenvalues and $e^{tA}$, variation of parameters, phase portraits and stability.
Most laws of nature couple several quantities at once: two competing species, the currents in an electrical network, the concentrations in a chain of chemical reactions, the position and velocity of a pendulum. Their mathematical form is a system $x'=f(t,x)$ for an unknown vector $x(t)\in\mathbb{R}^n$. Even a single equation of order $n$ is a system in disguise: the state of the pendulum $y''=-\sin y$ is not $y$ alone but the pair $(y,y')$. Written in vector notation, the whole theory of §1.1 — Picard iterates, Lipschitz conditions, maximal solutions, Gronwall's inequality — carries over almost word for word. For linear systems it gives far more: the solutions form an $n$-dimensional vector space, the determinant of a basis obeys an explicit formula (Liouville's formula), and when the coefficients are constant every solution is $e^{tA}x_0$, which can be computed from the eigenvalues and eigenvectors of $A$.
This section develops that theory and its geometry. For a $2\times2$ system the trace and the determinant alone decide whether the origin is a saddle, a node, a spiral or a centre and whether it is stable; linearisation carries the answer over to nonlinear systems near most equilibria. CSIR NET examines exactly these skills: computing $e^{tA}$ (above all with repeated or complex eigenvalues), evaluating Wronskians by Liouville's formula without solving, classifying critical points, deciding stability from eigenvalues, and solving non-homogeneous systems by variation of parameters. Chapter 2 rests on this section — a scalar linear equation of order $n$ is the special case of a companion matrix — and the small oscillations of Chapter 8 are linear systems with purely imaginary eigenvalues.
Notation. The independent variable is $t$; $x=(x_1,\dots,x_n)^{T}$ is a column vector in $\mathbb{R}^n$ (or $\mathbb{C}^n$). We fix a norm $\|\cdot\|$ on $\mathbb{R}^n$ (the Euclidean norm unless we say otherwise) and give an $n\times n$ matrix $A$ the induced norm $\|A\|=\max_{\|x\|=1}\|Ax\|$, so that $\|Ax\|\le\|A\|\,\|x\|$ and $\|AB\|\le\|A\|\,\|B\|$. Vector- and matrix-valued functions are differentiated and integrated entry by entry, and $\left\|\int_a^b v(s)\,ds\right\|\le\int_a^b\|v(s)\|\,ds$ for $a\le b$ (apply the triangle inequality to Riemann sums and pass to the limit).
Systems, Reduction of Order and Existence
Let $D\subseteq\mathbb{R}\times\mathbb{R}^n$ and let $f=(f_1,\dots,f_n)^T:D\to\mathbb{R}^n$ be continuous. The first-order system $x'=f(t,x)$ means $x_i'=f_i(t,x_1,\dots,x_n)$, $i=1,\dots,n$; with a condition $x(t_0)=x_0$, $(t_0,x_0)\in D$, it is an initial value problem. A solution on an interval $I\ni t_0$ is a differentiable $\varphi:I\to\mathbb{R}^n$ with $(t,\varphi(t))\in D$, $\varphi'(t)=f(t,\varphi(t))$ for all $t\in I$, and $\varphi(t_0)=x_0$.
$f$ is Lipschitz in $x$ on $D$ with constant $L$ if $\|f(t,x)-f(t,z)\|\le L\|x-z\|$ whenever $(t,x),(t,z)\in D$. The system is autonomous if $f$ does not depend on $t$, and linear if $f(t,x)=A(t)x+b(t)$ with an $n\times n$ matrix function $A$ and a vector function $b$; a linear system is homogeneous if $b\equiv0$.
All norms on $\mathbb{R}^n$ are equivalent, so whether $f$ is Lipschitz does not depend on the norm — only the value of the constant does.
Let $g$ be continuous on an open set $U\subseteq\mathbb{R}\times\mathbb{R}^n$. A function $y$ is a solution of
$$y^{(n)}=g\big(t,y,y',\dots,y^{(n-1)}\big),$$$y^{(k)}(t_0)=\eta_k$ ($0\le k\le n-1$), if and only if $x=(y,y',\dots,y^{(n-1)})^T$ is a solution of the system
$$\begin{aligned}x_k'&=x_{k+1}\quad(1\le k\le n-1),\\x_n'&=g(t,x_1,\dots,x_n),\end{aligned}$$with $x(t_0)=(\eta_0,\dots,\eta_{n-1})^T$, and every solution $x$ of the system has this form, with $y=x_1$. If $g$ is Lipschitz in its last $n$ variables with constant $L$, the right-hand side $F$ of the system is Lipschitz in $x$ (with constant $\max(1,L)$ in the max-norm). In particular the linear equation $y^{(n)}+a_{n-1}(t)y^{(n-1)}+\dots+a_0(t)y=r(t)$ becomes $x'=C(t)x+r(t)e_n$ with the companion matrix
$$C(t)=\begin{pmatrix}0&1&&\\&\ddots&\ddots&\\&&0&1\\-a_0&-a_1&\cdots&-a_{n-1}\end{pmatrix}.$$Proof. If $y$ solves the equation, the components $x_k=y^{(k-1)}$ satisfy $x_k'=y^{(k)}=x_{k+1}$ for $k\lt n$ and $x_n'=y^{(n)}=g(t,x)$, with the stated initial values. Conversely, let $x$ solve the system. The equations $x_{k+1}=x_k'$ give, by induction on $k$, that $x_1$ is $k$ times differentiable with $x_1^{(k)}=x_{k+1}$ for $k=1,\dots,n-1$. Since $x_n=x_1^{(n-1)}$ is differentiable, $x_1$ is $n$ times differentiable, and the last equation reads $x_1^{(n)}=g(t,x_1,x_1',\dots,x_1^{(n-1)})$. For the Lipschitz statement use the max-norm $\|x\|_\infty=\max_k|x_k|$: the first $n-1$ components of $F(t,x)-F(t,z)$ are $x_{k+1}-z_{k+1}$, of absolute value at most $\|x-z\|_\infty$, and the last is $g(t,x)-g(t,z)$, of absolute value at most $L\|x-z\|_\infty$. ∎
The converse operation — eliminating unknowns from a system to obtain one higher-order equation — often works too (Exercise I.2), but not always in a way that determines every component: in the uncoupled system $x_1'=x_1$, $x_2'=x_2$ no equation for $x_1$ alone says anything about $x_2$. Systems are the more general and the more natural object.
- (Picard–Lindelöf.) Let $R=\{(t,x):|t-t_0|\le a,\ \|x-x_0\|\le b\}$, let $f$ be continuous on $R$ with $\|f\|\le M$ there, and Lipschitz in $x$ on $R$ with constant $L$. With $h=\min(a,b/M)$, the IVP $x'=f(t,x)$, $x(t_0)=x_0$ has exactly one solution on $[t_0-h,t_0+h]$; it is the uniform limit of the Picard iterates $\varphi_0=x_0$, $\varphi_{k+1}(t)=x_0+\int_{t_0}^tf(s,\varphi_k(s))\,ds$.
- (A sufficient condition.) If $D$ is convex in $x$ (it contains the segment joining $(t,x)$ and $(t,z)$ whenever it contains both points), the partial derivatives $\partial f_i/\partial x_j$ exist and are continuous on $D$, and the Jacobian matrix $J(t,x)=\left(\partial f_i/\partial x_j\right)$ satisfies $\|J(t,x)\|\le L$ on $D$, then $f$ is Lipschitz in $x$ on $D$ with constant $L$. If $f$ is $C^1$ on an open set $D$, it is locally Lipschitz in $x$.
- (Maximal solutions.) If $D$ is open and $f$ is continuous and locally Lipschitz in $x$, every IVP has a unique maximal solution, defined on an open interval $(\omega_-,\omega_+)$, and $(t,x(t))$ leaves every compact subset of $D$ as $t\to\omega_\pm$. For $D=(\alpha,\beta)\times\mathbb{R}^n$: either $\omega_+=\beta$ or $\|x(t)\|\to\infty$ as $t\to\omega_+^-$.
- (Global existence.) If $f$ is continuous on $[\alpha,\beta]\times\mathbb{R}^n$ and Lipschitz in $x$ there, every IVP with $t_0\in[\alpha,\beta]$ has a unique solution on all of $[\alpha,\beta]$. In particular, if $A(t)$ and $b(t)$ are continuous on an interval $J$, then for every $t_0\in J$ and $x_0\in\mathbb{R}^n$ the linear IVP $x'=A(t)x+b(t)$, $x(t_0)=x_0$ has a unique solution, and it is defined on the whole of $J$.
Proof. (1), (3) and the first half of (4) are Theorems 1.3, 1.5 and 1.6 with $|\cdot|$ replaced by $\|\cdot\|$. Those proofs used only the triangle inequality, the estimate $\left\|\int v\right\|\le\int\|v\|$, the Lipschitz condition, the Weierstrass M-test (applied component by component) and compactness — all valid for vector functions. For instance the key estimate of Step 2 becomes $$\|\varphi_{k+1}(t)-\varphi_k(t)\|\le\left|\int_{t_0}^t\big\|f(s,\varphi_k(s))-f(s,\varphi_{k-1}(s))\big\|\,ds\right|\le\frac{ML^k|t-t_0|^{k+1}}{(k+1)!},$$ word for word as before. Gronwall's inequality and the continuous-dependence estimate (Theorems 1.7, 1.8) also hold verbatim with norms.
(2) Fix $t$ and two points $x,z$ with $(t,x),(t,z)\in D$, and put $g(\sigma)=f\big(t,z+\sigma(x-z)\big)$ for $\sigma\in[0,1]$; the segment lies in $D$ by convexity. By the chain rule $g'(\sigma)=J\big(t,z+\sigma(x-z)\big)(x-z)$, which is continuous, so $$f(t,x)-f(t,z)=g(1)-g(0)=\int_0^1J\big(t,z+\sigma(x-z)\big)(x-z)\,d\sigma,$$ and therefore $\|f(t,x)-f(t,z)\|\le\int_0^1L\|x-z\|\,d\sigma=L\|x-z\|$. (We integrate instead of using the mean value theorem, which is false for vector-valued functions — see the pitfalls below.) Locally, a closed box around any point of an open $D$ is compact and convex, and the continuous function $\|J\|$ is bounded on it.
(4), linear case: the difference of the right-hand sides at $x$ and at $z$ is $A(t)(x-z)$, and $\|A(t)(x-z)\|\le\|A(t)\|\,\|x-z\|$. On every compact interval $K\subseteq J$ the continuous function $t\mapsto\|A(t)\|$ is bounded by some $L_K$, so the right-hand side is Lipschitz on $K\times\mathbb{R}^n$ and the first half of (4) gives a unique solution on $K$. Exhaust $J$ by an increasing sequence of compact intervals containing $t_0$; by uniqueness the solutions agree on overlaps and define one solution on $J$. ∎
Combining Theorem 1.17 with part (4): a linear equation of order $n$ with coefficients continuous on $J$ has, for any $n$ prescribed values $y(t_0),\dots,y^{(n-1)}(t_0)$, exactly one solution, and it lives on all of $J$. This is the starting point of Chapter 2.
Linear Systems: Fundamental Matrices and Liouville's Formula
From now on $A(t)$ is an $n\times n$ matrix function, continuous on an interval $J$ (real or complex entries), and we study the homogeneous system $$x'=A(t)x\qquad(t\in J).\tag{H}$$ By Theorem 1.18(4) every IVP for (H) has a unique solution on all of $J$.
Proof. If $\varphi,\psi\in S$ and $c,d$ are scalars, then $(c\varphi+d\psi)'=cA\varphi+dA\psi=A(c\varphi+d\psi)$, so $S$ is a vector space, and evaluation at $t_0$ is linear. It is injective: if $\varphi(t_0)=0$, then $\varphi$ and the zero function solve the same IVP, so $\varphi\equiv0$ by uniqueness. It is surjective: for every $x_0$ the IVP with $x(t_0)=x_0$ has a solution on $J$. Hence $\dim S=n$. Finally, by injectivity, $\sum c_i\varphi_i\equiv0$ on $J$ if and only if $\sum c_i\varphi_i(t_0)=0$, so the linear relations among the functions and among their values at $t_0$ are the same. ∎
Proof. Let $r_1,\dots,r_n$ be the rows of $\Phi(t)$. The determinant is linear in each row separately, so by the product rule its derivative is the sum of the $n$ determinants obtained by differentiating one row at a time: $$(\det\Phi)'=\sum_{i=1}^n\det\big(\Phi\text{ with }r_i\text{ replaced by }r_i'\big).$$ From $\Phi'=A\Phi$, the $i$-th row of $\Phi'$ is $r_i'=\sum_{k}a_{ik}r_k$. By linearity in row $i$, the $i$-th determinant equals $\sum_ka_{ik}\det\big(\Phi\text{ with }r_i\text{ replaced by }r_k\big)$. For $k\ne i$ that matrix has two equal rows and determinant $0$; for $k=i$ it is $\Phi$ itself. So the $i$-th term is $a_{ii}\det\Phi$, and $$W'=\Big(\sum_ia_{ii}(t)\Big)W=\operatorname{tr}A(t)\,W.$$ This scalar linear equation has the unique solution $W(t)=W(t_0)\exp\int_{t_0}^t\operatorname{tr}A$ (Theorem 1.10(2)). The exponential never vanishes, so $W$ vanishes at one point iff $W(t_0)=0$ iff $W\equiv0$; and by Theorem 1.19, $W(t_0)\ne0$ iff the columns are independent. ∎
Geometric meaning. $|W(t)|$ is the volume of the parallelepiped spanned by the solution vectors $\varphi_1(t),\dots,\varphi_n(t)$; the flow of (H) changes volumes at the relative rate $\operatorname{tr}A(t)$, the divergence of the vector field $x\mapsto A(t)x$. Interactive 1 below shows this for $n=2$. Abel's formula is the special case of a companion matrix: for $y''+p(t)y'+q(t)y=0$ the matrix is $\begin{pmatrix}0&1\\-q&-p\end{pmatrix}$ with trace $-p$, and the Wronskian $W=y_1y_2'-y_1'y_2$ of two solutions satisfies $W(t)=W(t_0)\exp\left(-\int_{t_0}^tp\right)$; for order $n$ the trace is $-a_{n-1}$.
- A matrix function $\Psi$ is a fundamental matrix of (H) if and only if $\Psi(t)=\Phi(t)C$ for a constant nonsingular matrix $C$.
- Every solution of (H) is $x(t)=\Phi(t)c$ for a unique $c\in\mathbb{R}^n$, and the solution with $x(t_0)=x_0$ is $$x(t)=\Phi(t)\Phi(t_0)^{-1}x_0,\qquad\text{i.e.}\qquad\Phi(t,t_0)=\Phi(t)\Phi(t_0)^{-1}.$$
- $\Phi(t,s)\Phi(s,r)=\Phi(t,r)$ and $\Phi(t,s)^{-1}=\Phi(s,t)$ for all $t,s,r\in J$.
Proof. (1) If $\Psi=\Phi C$, then $\Psi'=\Phi'C=A\Phi C=A\Psi$ and $\det\Psi=\det\Phi\det C\ne0$. Conversely, let $\Psi$ be fundamental. By Liouville's formula $\Phi(t)$ is invertible for every $t$, and $\Phi^{-1}$ is differentiable (its entries are cofactors divided by $\det\Phi$). Put $C(t)=\Phi(t)^{-1}\Psi(t)$. Differentiating $\Phi C=\Psi$ gives $\Phi'C+\Phi C'=\Psi'$, i.e. $A\Phi C+\Phi C'=A\Psi=A\Phi C$, so $\Phi C'=0$ and $C'=0$: $C$ is constant, and nonsingular because $\det\Psi\ne0$. (2) $\Phi c=\sum c_i\varphi_i$, and the columns form a basis of $S$ (Theorem 1.19). The condition $\Phi(t_0)c=x_0$ gives $c=\Phi(t_0)^{-1}x_0$. (3) For fixed $s$ and $r$, both $t\mapsto\Phi(t,s)\Phi(s,r)$ and $t\mapsto\Phi(t,r)$ are matrix solutions with the value $\Phi(s,r)$ at $t=s$; by uniqueness (column by column) they coincide. Taking $r=t$ gives $\Phi(t,s)\Phi(s,t)=I$. ∎
Constant Coefficients: Eigenvalues and the Matrix Exponential
For the scalar equation $x'=ax$ the solution is $e^{ta}x_0$. For a constant matrix $A$ the same formula works once the exponential of a matrix is defined by its power series.
- The series converges absolutely for every $M$, with $\|e^M\|\le e^{\|M\|}$, and the series for $e^{tA}$ converges uniformly on bounded $t$-intervals.
- $\dfrac{d}{dt}e^{tA}=Ae^{tA}=e^{tA}A$. Hence $e^{tA}$ is the principal fundamental matrix at $0$ of $x'=Ax$, and the solution with $x(t_0)=x_0$ is $x(t)=e^{(t-t_0)A}x_0$.
- If $AB=BA$, then $e^{A+B}=e^Ae^B$. In particular $e^{(s+t)A}=e^{sA}e^{tA}$, $\left(e^{tA}\right)^{-1}=e^{-tA}$, and $e^{t(\lambda I+N)}=e^{\lambda t}e^{tN}$.
- $e^{P^{-1}AP}=P^{-1}e^AP$, and $e^{\operatorname{diag}(\lambda_1,\dots,\lambda_n)}=\operatorname{diag}(e^{\lambda_1},\dots,e^{\lambda_n})$.
- $\det e^{tA}=e^{t\operatorname{tr}A}$; in particular $e^{tA}$ is always invertible.
Proof. (1) $\|M^k/k!\|\le\|M\|^k/k!$ and $\sum_k\|M\|^k/k!=e^{\|M\|}$; since every entry of a matrix is bounded by its norm, every entry series converges absolutely. For $|t|\le T$, $\|t^kA^k/k!\|\le(T\|A\|)^k/k!$, a convergent numerical series, so the Weierstrass M-test gives uniform convergence. (2) The termwise derivative $\sum_{k\ge1}t^{k-1}A^k/(k-1)!=A\sum_{j\ge0}t^jA^j/j!$ is again a uniformly convergent series on bounded intervals, so termwise differentiation is legitimate and gives $Ae^{tA}$; $A$ commutes with every partial sum, hence with $e^{tA}$. Since $e^{0A}=I$, $e^{tA}$ is the principal fundamental matrix at $0$, and $x(t)=e^{(t-t_0)A}x_0$ satisfies $x'=Ax$, $x(t_0)=x_0$; uniqueness does the rest.
(3) Let $Y(t)=e^{t(A+B)}$ and $Z(t)=e^{tA}e^{tB}$. Because $B$ commutes with $A$, it commutes with every partial sum of $e^{tA}$ and hence with $e^{tA}$. So $Z'=Ae^{tA}e^{tB}+e^{tA}Be^{tB}=(A+B)Z$, and $Y'=(A+B)Y$; both equal $I$ at $t=0$, so $Y=Z$ by uniqueness, and $t=1$ gives the claim. The special cases use that $sA$ and $tA$, $A$ and $-A$, $\lambda I$ and $N$ commute. (4) $(P^{-1}AP)^k=P^{-1}A^kP$; sum the series. A diagonal matrix has diagonal powers. (5) Liouville's formula with $\Phi=e^{tA}$, $\det\Phi(0)=1$ and constant trace. ∎
- If $Av=\lambda v$ with $v\ne0$ ($\lambda,v$ possibly complex), then $e^{tA}v=e^{\lambda t}v$, and $x(t)=e^{\lambda t}v$ is a solution of $x'=Ax$.
- If $A$ has $n$ linearly independent eigenvectors $v_1,\dots,v_n$ with eigenvalues $\lambda_1,\dots,\lambda_n$ (for instance if the eigenvalues are distinct), then $\Phi(t)=\big[e^{\lambda_1t}v_1\ \cdots\ e^{\lambda_nt}v_n\big]$ is a fundamental matrix and $$e^{tA}=P\operatorname{diag}\left(e^{\lambda_1t},\dots,e^{\lambda_nt}\right)P^{-1},\qquad P=[v_1\ \cdots\ v_n].$$
- If $A$ is real and $\lambda=\alpha+i\beta$ with $\beta\ne0$ is an eigenvalue with eigenvector $v=a+ib$ ($a,b\in\mathbb{R}^n$), then $$\operatorname{Re}\left(e^{\lambda t}v\right)=e^{\alpha t}(a\cos\beta t-b\sin\beta t),\qquad\operatorname{Im}\left(e^{\lambda t}v\right)=e^{\alpha t}(a\sin\beta t+b\cos\beta t)$$ are two linearly independent real solutions; together they replace the complex pair belonging to $\lambda$ and $\bar\lambda$.
Proof. (1) $A^kv=\lambda^kv$, so $e^{tA}v=\sum_k\frac{(t\lambda)^k}{k!}v=e^{\lambda t}v$, and $(e^{\lambda t}v)'=\lambda e^{\lambda t}v=e^{\lambda t}Av=A(e^{\lambda t}v)$. (2) The columns are solutions and $\Phi(0)=P$ is invertible, so $\Phi$ is fundamental (Theorem 1.20). Since $\Phi(t)=P\operatorname{diag}(e^{\lambda_it})$, Theorem 1.21(2) gives $e^{tA}=\Phi(t)\Phi(0)^{-1}=P\operatorname{diag}(e^{\lambda_it})P^{-1}$. (3) If $z$ is a complex solution of $x'=Ax$ with $A$ real, then taking real and imaginary parts of $z'=Az$ shows that $\operatorname{Re}z$ and $\operatorname{Im}z$ are real solutions. Their values at $t=0$ are $a$ and $b$. If $a,b$ were linearly dependent over $\mathbb{R}$, say $b=ca$ (the case $a=cb$ is the same), then $v=(1+ic)a$ and $a\ne0$ would be a real eigenvector for the non-real eigenvalue $\lambda$; but then $Aa$ is real while $\lambda a$ has imaginary part $\beta a\ne0$ — impossible. So the two solutions are independent (Theorem 1.19). ∎
When $A$ has too few eigenvectors, polynomial factors $t^k$ appear. The key is that $\lambda I$ commutes with everything.
- The generalised eigenspace $G_\lambda=\ker(A-\lambda I)^m$ has dimension $m$ and is invariant under $A$, and $\mathbb{C}^n$ is the direct sum of the $G_\lambda$ over the distinct eigenvalues (primary decomposition).
- For every $v\in G_\lambda$ the series terminates: $$e^{tA}v=e^{\lambda t}\sum_{k=0}^{m-1}\frac{t^k}{k!}(A-\lambda I)^kv.$$
- Hence every solution of $x'=Ax$ is a finite sum of terms $t^ke^{\lambda t}u$ with $\lambda$ an eigenvalue, $u$ a constant vector and $0\le k\le m_\lambda-1$; terms with $k\ge1$ occur if and only if $A$ is not diagonalisable.
- ($2\times2$ formulas.) Let $n=2$ with eigenvalues $\lambda_1,\lambda_2$. Then $$\begin{aligned}\lambda_1=\lambda_2=\lambda:&\quad e^{tA}=e^{\lambda t}\big(I+t(A-\lambda I)\big);\\ \lambda_1\ne\lambda_2:&\quad e^{tA}=\frac{e^{\lambda_1t}(A-\lambda_2I)-e^{\lambda_2t}(A-\lambda_1I)}{\lambda_1-\lambda_2};\\ \lambda=\alpha\pm i\beta,\ \beta\gt0:&\quad e^{tA}=e^{\alpha t}\left(\cos\beta t\,I+\frac{\sin\beta t}{\beta}(A-\alpha I)\right).\end{aligned}$$
Proof. (1) is the primary decomposition theorem of linear algebra (treated in the CSIR NET Unit 1 course, in its chapter on canonical forms); we use it as stated. (2) $\lambda I$ commutes with $A-\lambda I$, so by Theorem 1.22(3) $e^{tA}=e^{t\lambda I}e^{t(A-\lambda I)}=e^{\lambda t}e^{t(A-\lambda I)}$. For $v\in G_\lambda$, $(A-\lambda I)^kv=0$ for $k\ge m$, so the series for $e^{t(A-\lambda I)}v$ stops after the term $k=m-1$. (3) Decompose $x_0=\sum_\lambda v_\lambda$ with $v_\lambda\in G_\lambda$ by (1), and apply (2) to each piece of $e^{tA}x_0=\sum_\lambda e^{tA}v_\lambda$. If $A$ is diagonalisable, each $G_\lambda$ is the eigenspace, $(A-\lambda I)v_\lambda=0$, and only $k=0$ occurs. If not, some $G_\lambda$ is larger than $\ker(A-\lambda I)$; choosing $v\in G_\lambda$ with $(A-\lambda I)v\ne0$, the solution $e^{tA}v$ contains the term $te^{\lambda t}(A-\lambda I)v\ne0$ (it is not cancelled, because $e^{-\lambda t}e^{tA}v$ is a polynomial in $t$ with that coefficient of $t$).
(4) By the Cayley–Hamilton theorem $(A-\lambda_1I)(A-\lambda_2I)=0$. Double eigenvalue: $(A-\lambda I)^2=0$, so every $v$ lies in $G_\lambda$ with $m=2$, and (2) gives the formula. Distinct eigenvalues: put $P_1=\frac{A-\lambda_2I}{\lambda_1-\lambda_2}$, $P_2=\frac{A-\lambda_1I}{\lambda_2-\lambda_1}$. Then $P_1+P_2=I$ and, by Cayley–Hamilton, $(A-\lambda_1I)P_1=0$, i.e. $AP_1=\lambda_1P_1$; likewise $AP_2=\lambda_2P_2$. By induction $A^k=\lambda_1^kP_1+\lambda_2^kP_2$, so $e^{tA}=e^{\lambda_1t}P_1+e^{\lambda_2t}P_2$, which is the formula (valid also for complex $\lambda_1\ne\lambda_2$). Complex eigenvalues: the characteristic polynomial is $(\mu-\alpha)^2+\beta^2$, so $N=A-\alpha I$ satisfies $N^2=-\beta^2I$. Then $N^{2j}=(-1)^j\beta^{2j}I$, $N^{2j+1}=(-1)^j\beta^{2j}N$, and $$e^{tN}=\sum_j\frac{(-1)^j(\beta t)^{2j}}{(2j)!}I+\sum_j\frac{(-1)^j\beta^{2j}t^{2j+1}}{(2j+1)!}N=\cos\beta t\,I+\frac{\sin\beta t}{\beta}N;$$ multiply by $e^{\alpha t}$ (Theorem 1.22(3)). ∎
- Eigenvalues first. For $n=2$: $\lambda^2-\tau\lambda+\Delta=0$ with $\tau=\operatorname{tr}A$, $\Delta=\det A$.
- $2\times2$: use Theorem 1.24(4) directly — no eigenvectors needed. Check with $e^{0A}=I$ and $\det e^{tA}=e^{t\tau}$.
- Diagonalisable: for an IVP, expand $x_0=\sum c_iv_i$ in eigenvectors; then $x(t)=\sum c_ie^{\lambda_it}v_i$. For complex pairs take real and imaginary parts of one complex solution.
- Not diagonalisable: compute $G_\lambda=\ker(A-\lambda I)^{m}$, split $x_0$ into its components in the $G_\lambda$, and apply the finite sum of Theorem 1.24(2) to each.
- Interpolation alternative. $e^{tA}=r(A)$, where $r$ is the polynomial of degree $\lt n$ with $r^{(j)}(\lambda)=\frac{d^j}{d\lambda^j}e^{\lambda t}$ for $0\le j\lt m_\lambda$ at every eigenvalue. (By Theorem 1.24(2) and Taylor's formula both sides act in the same way on every $G_\lambda$.) For $2\times2$ with distinct eigenvalues: $e^{tA}=\alpha_0(t)I+\alpha_1(t)A$ with $\alpha_0+\alpha_1\lambda_i=e^{\lambda_it}$.
- Nilpotent pieces ($N^k=0$): $e^{tN}=I+tN+\dots+\frac{t^{k-1}}{(k-1)!}N^{k-1}$, and $e^{t(\lambda I+N)}=e^{\lambda t}e^{tN}$.
Non-Homogeneous Systems: Variation of Parameters
For $x'=A(t)x+b(t)$ the idea of Lagrange is to let the constant vector $c$ in the general solution $\Phi(t)c$ of the homogeneous system vary with $t$.
- The solutions of $x'=A(t)x+b(t)$ are exactly $x=x_p+\Phi(t)c$, $c\in\mathbb{R}^n$, where $x_p$ is any one solution: they form an affine space of dimension $n$.
- The solution with $x(t_0)=x_0$ is $$x(t)=\Phi(t)\Phi(t_0)^{-1}x_0+\Phi(t)\int_{t_0}^{t}\Phi(s)^{-1}b(s)\,ds.$$
- For constant $A$: $$x(t)=e^{(t-t_0)A}x_0+\int_{t_0}^{t}e^{(t-s)A}\,b(s)\,ds.$$
Proof. (1) The difference of two solutions solves the homogeneous system, so it equals $\Phi(t)c$ for some $c$ (Theorem 1.21(2)); conversely $x_p+\Phi c$ is a solution. (2) Look for a solution $x=\Phi(t)c(t)$. Then $x'=\Phi'c+\Phi c'=A\Phi c+\Phi c'=Ax+\Phi c'$, so $x$ solves the system if and only if $\Phi(t)c'(t)=b(t)$, i.e. $c'=\Phi^{-1}b$ — here we use that $\Phi(t)$ is invertible for every $t$ (Liouville). The function $\Phi^{-1}b$ is continuous, so $c(t)=\Phi(t_0)^{-1}x_0+\int_{t_0}^t\Phi(s)^{-1}b(s)\,ds$ is $C^1$, and the computation shows that $x=\Phi c$ solves the IVP; it is the only solution by Theorem 1.18(4). (3) Take $\Phi(t)=e^{tA}$; then $\Phi(t)\Phi(s)^{-1}=e^{tA}e^{-sA}=e^{(t-s)A}$. ∎
The formula has a transparent meaning (Duhamel's principle): the input $b(s)\,ds$ injected at time $s$ is carried forward to time $t$ by the homogeneous flow $e^{(t-s)A}$, and the response is the superposition of all these contributions. The same kernel $e^{(t-s)A}$ is the Green's function for initial value problems in §2.4, and the formula is the main tool in the proof of Theorem 1.28 below.
Phase Portraits and Stability
Let $f:\Omega\to\mathbb{R}^n$ be locally Lipschitz on an open set $\Omega$ and consider the autonomous system $x'=f(x)$. The trajectory (orbit) through $x_0$ is the curve traced by the maximal solution with $x(0)=x_0$; the phase portrait is the family of all trajectories, with arrows pointing in the direction of increasing $t$. A point $x^*$ with $f(x^*)=0$ is an equilibrium (critical point): $x(t)\equiv x^*$ is a solution. The equilibrium $x^*$ is
- stable (in the sense of Lyapunov) if for every $\varepsilon\gt0$ there is $\delta\gt0$ such that $\|x_0-x^*\|\lt\delta$ implies that the solution exists on $[0,\infty)$ and $\|x(t)-x^*\|\lt\varepsilon$ for all $t\ge0$;
- asymptotically stable if it is stable and 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.
For a linear system $x'=Ax$ the origin is always an equilibrium; it is isolated iff $\det A\ne0$. By linearity (scale the initial value), the origin of $x'=Ax$ is stable iff every solution is bounded on $[0,\infty)$, and asymptotically stable iff every solution tends to $0$.
Trajectories of an autonomous system never cross: if $x(t_1)=z(t_2)$, then $t\mapsto x(t+t_1-t_2)$ and $z$ solve the same IVP at $t_2$, so they coincide — the two curves are the same trajectory traversed with a time shift. For non-autonomous systems this fails, which is one reason the phase plane is reserved for autonomous systems.
- a saddle if exactly two trajectories tend to it as $t\to+\infty$ (along one line), two tend to it as $t\to-\infty$ (along another line), and all others come close and leave again;
- a node if all trajectories tend to it as $t\to+\infty$ (stable node) or all as $t\to-\infty$ (unstable node), each with a limiting tangent direction; a star (proper node) if the trajectories are rays, so every direction occurs, and a degenerate (improper) node if all trajectories are tangent to a single line;
- a spiral point (focus) if all trajectories wind around it infinitely often while tending to it as $t\to+\infty$ (stable) or as $t\to-\infty$ (unstable);
- a centre if all other trajectories are closed curves surrounding it (periodic solutions).
Let $A=\begin{pmatrix}a&b\\c&d\end{pmatrix}$ be real, $\tau=\operatorname{tr}A$, $\Delta=\det A\ne0$ and $D=\tau^2-4\Delta$, so that the eigenvalues are $\lambda=\tfrac12\left(\tau\pm\sqrt D\right)$. The origin of $x'=Ax$ is:
| condition | eigenvalues | type | stability |
|---|---|---|---|
| $\Delta\lt0$ | real, opposite signs | saddle | unstable |
| $\Delta\gt0$, $D\gt0$ | real, distinct, sign of $\tau$ | node | asympt. stable iff $\tau\lt0$ |
| $\Delta\gt0$, $D=0$ | $\lambda=\tau/2$ double | star if $A=\lambda I$, else degenerate node | asympt. stable iff $\tau\lt0$ |
| $\Delta\gt0$, $D\lt0$, $\tau\ne0$ | $\tfrac\tau2\pm i\beta$ | spiral | asympt. stable iff $\tau\lt0$ |
| $\Delta\gt0$, $\tau=0$ | $\pm i\sqrt\Delta$ | centre (ellipses) | stable, not asymptotically |
For a spiral or centre the rotation is counterclockwise if $c\gt0$ and clockwise if $c\lt0$. In particular the origin is asymptotically stable if and only if $\tau\lt0$ and $\Delta\gt0$.
Proof. Recall $\Delta=\lambda_1\lambda_2$ and $\tau=\lambda_1+\lambda_2$. In each case we pass to coordinates $y=P^{-1}x$ in which $A$ is simple; a linear change of coordinates maps lines to lines, rays to rays and circles to ellipses, and preserves boundedness and convergence to $0$.
Real distinct eigenvalues ($D\gt0$). With eigenvectors $v_1,v_2$ and $P=[v_1\ v_2]$ we get $y_1=c_1e^{\lambda_1t}$, $y_2=c_2e^{\lambda_2t}$. If $\Delta\lt0$, say $\lambda_1\gt0\gt\lambda_2$: the quantity $|y_1|^{-\lambda_2}|y_2|^{\lambda_1}$ is constant along solutions, so the trajectories off the axes are hyperbola-like curves, the half-axes of $y_2$ are the two trajectories entering the origin and those of $y_1$ the two leaving it: a saddle. It is unstable, since $\varepsilon e^{\lambda_1t}v_1$ starts arbitrarily close and is unbounded. If $\Delta\gt0$ the eigenvalues have the common sign of $\tau$; say $|\lambda_1|\lt|\lambda_2|$. Then $|y_2|=C|y_1|^{\lambda_2/\lambda_1}$ with exponent $\lambda_2/\lambda_1\gt1$, so all trajectories with $c_1\ne0$ are tangent at the origin to the slow direction $v_1$: a node, stable iff both eigenvalues are negative iff $\tau\lt0$.
Double eigenvalue ($D=0$, $\lambda=\tau/2\ne0$). If $A=\lambda I$, then $x=e^{\lambda t}x_0$: rays, a star. Otherwise $A-\lambda I\ne0$ but $(A-\lambda I)^2=0$ (Cayley–Hamilton); pick $v$ with $w=(A-\lambda I)v\ne0$, an eigenvector. In the basis $(w,v)$, $y=e^{\lambda t}(c_1+c_2t,\ c_2)$, whose direction tends to the $w$-axis as $t\to\pm\infty$: all trajectories are tangent to one line, a degenerate node, decaying iff $\lambda\lt0$.
Complex eigenvalues ($D\lt0$, $\lambda=\alpha\pm i\beta$, $\alpha=\tau/2$, $\beta\gt0$). With $v=a+ib$ an eigenvector for $\alpha+i\beta$, comparing real and imaginary parts of $Av=\lambda v$ gives $Aa=\alpha a-\beta b$ and $Ab=\beta a+\alpha b$; so with $P=[a\ b]$, $y'=\begin{pmatrix}\alpha&\beta\\-\beta&\alpha\end{pmatrix}y$. In polar coordinates of $y$ this reads $r'=\alpha r$, $\theta'=-\beta$: $r=r_0e^{\alpha t}$, $\theta=\theta_0-\beta t$, logarithmic spirals for $\alpha\ne0$ and circles for $\alpha=0$. So we get a spiral, stable iff $\alpha\lt0$, or, for $\tau=0$, a centre; at a centre $\|y(t)\|$ is constant, hence $\|x(t)\|\le\|P\|\,\|P^{-1}\|\,\|x_0\|$ (stable) but $x(t)\not\to0$. For the sense of rotation compute the angular velocity in the $x$-plane: $$\frac{d}{dt}\arg x=\frac{x_1x_2'-x_2x_1'}{x_1^2+x_2^2}=\frac{c\,x_1^2+(d-a)x_1x_2-b\,x_2^2}{x_1^2+x_2^2}.$$ The quadratic form in the numerator has discriminant $(d-a)^2+4bc=D\lt0$, so it is definite, with the sign of $c$. The last statement of the theorem collects the rows with $\tau\lt0$, $\Delta\gt0$. ∎
- The origin is asymptotically stable if and only if every eigenvalue satisfies $\operatorname{Re}\lambda\lt0$. In that case there are constants $K\ge1$, $\mu\gt0$ with $\|e^{tA}\|\le Ke^{-\mu t}$ for all $t\ge0$.
- The origin is stable if and only if every eigenvalue satisfies $\operatorname{Re}\lambda\le0$ and every eigenvalue with $\operatorname{Re}\lambda=0$ is semisimple (its algebraic and geometric multiplicities are equal).
- Otherwise it is unstable; in particular it is unstable as soon as one eigenvalue has $\operatorname{Re}\lambda\gt0$.
Proof. For an eigenvalue $\lambda$ let $\nu_\lambda$ be the least $j$ with $\ker(A-\lambda I)^j=G_\lambda$; then $\nu_\lambda=1$ exactly when $\ker(A-\lambda I)=G_\lambda$, i.e. when $\lambda$ is semisimple. By Theorem 1.24, applied to the columns $e^{tA}e_j$ with $e_j$ split along the $G_\lambda$, every entry of $e^{tA}$ is a finite sum of terms $\gamma\,t^ke^{\lambda t}$ with $0\le k\lt\nu_\lambda$, and $|t^ke^{\lambda t}|=t^ke^{(\operatorname{Re}\lambda)t}$.
(1) If all $\operatorname{Re}\lambda\lt0$, fix $0\lt\mu\lt\min_\lambda(-\operatorname{Re}\lambda)$. Each $t^ke^{(\operatorname{Re}\lambda+\mu)t}$ tends to $0$ as $t\to\infty$, hence is bounded on $[0,\infty)$, so $e^{\mu t}\|e^{tA}\|\le K$ there. Then $\|x(t)\|\le Ke^{-\mu t}\|x_0\|$: stability (take $\delta=\varepsilon/K$) and convergence to $0$. Conversely, let some eigenvalue $\lambda=\alpha+i\beta$ have $\alpha\ge0$, with eigenvector $v=a+ib$. For real $\lambda$ the solution $\varepsilon e^{\lambda t}v$ does not tend to $0$. For $\beta\ne0$ the real solution $\varepsilon e^{\alpha t}(a\cos\beta t-b\sin\beta t)$ equals $\varepsilon e^{\alpha t}a$ at the times $t=2\pi k/\beta$, and $a\ne0$ (Theorem 1.23(3)), so it does not tend to $0$ either. Since $\varepsilon$ is arbitrary, the origin is not asymptotically stable.
(2) If the conditions hold, terms with $\operatorname{Re}\lambda\lt0$ are bounded on $[0,\infty)$ as in (1), and for $\operatorname{Re}\lambda=0$ only $k=0$ occurs, with $|e^{\lambda t}|=1$. So $\|e^{tA}\|\le K$ on $[0,\infty)$ and $\|x(t)\|\le K\|x_0\|$: stable. Conversely, if $\operatorname{Re}\lambda\gt0$, the solutions just constructed are unbounded. If $\operatorname{Re}\lambda=0$ and $\lambda$ is not semisimple, choose $v\in\ker(A-\lambda I)^2\setminus\ker(A-\lambda I)$ and put $w=(A-\lambda I)v\ne0$; then $e^{tA}v=e^{\lambda t}(v+tw)$ has norm at least $t\|w\|-\|v\|\to\infty$. Its real part or its imaginary part is then an unbounded real solution (if both were bounded, so would be their combination), and multiplying by $\varepsilon$ gives unbounded solutions starting arbitrarily close to $0$. In both cases the origin is unstable, which also proves (3). ∎
For a nonlinear autonomous system the eigenvalues of the Jacobian matrix at an equilibrium usually decide the matter.
- If every eigenvalue of $A$ has negative real part, $x^*$ is an asymptotically stable equilibrium of $x'=f(x)$.
- If some eigenvalue of $A$ has positive real part, $x^*$ is unstable.
- (Statement only.) In the plane, if $f$ is $C^2$ and $x^*$ is hyperbolic (no eigenvalue of $A$ has zero real part), then $x^*$ is a saddle, a node or a spiral of $x'=f(x)$ exactly when it is one for $y'=Ay$ (Hartman–Grobman theorem and its planar refinement). If $A$ has a centre, the linearisation decides nothing.
Proof of (1). Translate so that $x^*=0$ and write $f(x)=Ax+g(x)$. Differentiability of $f$ at $0$ means $\|g(x)\|/\|x\|\to0$: for every $\varepsilon\gt0$ there is $r\gt0$ with $\|g(x)\|\le\varepsilon\|x\|$ for $\|x\|\le r$. By Theorem 1.27(1), $\|e^{tA}\|\le Ke^{-\mu t}$ for $t\ge0$; choose $\varepsilon=\mu/(2K)$ and the corresponding $r$. A solution $x$ of $x'=Ax+g(x)$ is also a solution of the linear system with forcing $b(t)=g(x(t))$, so by Theorem 1.25(3) $$x(t)=e^{tA}x_0+\int_0^te^{(t-s)A}g\big(x(s)\big)\,ds.$$ As long as $\|x(s)\|\le r$ on $[0,t]$, this gives $\|x(t)\|\le Ke^{-\mu t}\|x_0\|+K\varepsilon\int_0^te^{-\mu(t-s)}\|x(s)\|\,ds$. Multiply by $e^{\mu t}$ and put $u(t)=e^{\mu t}\|x(t)\|$: $$u(t)\le K\|x_0\|+K\varepsilon\int_0^tu(s)\,ds,$$ so Gronwall's inequality (Theorem 1.7) yields $u(t)\le K\|x_0\|e^{K\varepsilon t}$, that is, $$\|x(t)\|\le K\|x_0\|\,e^{-(\mu-K\varepsilon)t}=K\|x_0\|\,e^{-\mu t/2}.$$ Now let $\|x_0\|\lt r/K$. If the solution ever reached the sphere $\|x\|=r$, at the first such time $t_1$ the estimate (valid on $[0,t_1]$) would give $\|x(t_1)\|\le K\|x_0\|\lt r$ — a contradiction. So the solution stays in the ball, exists for all $t\ge0$ (Theorem 1.18(3)), and decays exponentially: $x^*=0$ is asymptotically stable. We omit the proofs of (2) and (3). ∎
- $W\equiv0$ does not imply dependence for arbitrary vector functions. $(1,t)^T$ and $(t,t^2)^T$ have $W=t^2-t^2\equiv0$, yet neither is a constant multiple of the other. Liouville's dichotomy (Theorem 1.20) is about solutions of one system with continuous coefficients — so these two functions cannot both solve such a system on any interval.
- A Wronskian with an isolated zero signals a singular coefficient. $(t,1)^T,(t^2,2t)^T$ have $W=t^2$; they solve the system of Example 2, whose coefficients blow up at $t=0$.
- No mean value theorem for vector functions. $g(\sigma)=(\cos\sigma,\sin\sigma)$ has $g(2\pi)-g(0)=0$, but $g'(\sigma)\ne0$ for every $\sigma$. Lipschitz bounds come from the integral form (Theorem 1.18(2)).
- $e^{A+B}\ne e^Ae^B$ in general. For $A=\begin{pmatrix}0&1\\0&0\end{pmatrix}$, $B=\begin{pmatrix}0&0\\1&0\end{pmatrix}$: $e^Ae^B=\begin{pmatrix}2&1\\1&1\end{pmatrix}$, while $e^{A+B}=\begin{pmatrix}\cosh1&\sinh1\\\sinh1&\cosh1\end{pmatrix}$. Here $AB\ne BA$.
- $\exp\left(\int_0^tA(s)\,ds\right)$ is not a fundamental matrix in general. For $A(t)=\begin{pmatrix}0&1\\0&t\end{pmatrix}$ the principal fundamental matrix has $(1,2)$ entry $\int_0^te^{s^2/2}ds=t+\tfrac{t^3}6+\cdots$, whereas $\exp\int_0^tA$ has $(1,2)$ entry $\tfrac2t\left(e^{t^2/2}-1\right)=t+\tfrac{t^3}4+\cdots$. It works when $A(t)$ commutes with $\int_0^tA$, e.g. $A(t)=a(t)M$.
- Frozen eigenvalues do not decide stability of $x'=A(t)x$. For $A(t)=\begin{pmatrix}-1+\frac32\cos^2t&1-\frac32\cos t\sin t\\-1-\frac32\sin t\cos t&-1+\frac32\sin^2t\end{pmatrix}$ the eigenvalues are $\tfrac{-1\pm i\sqrt7}{4}$ for every $t$, yet $x(t)=e^{t/2}(-\cos t,\sin t)^T$ is a solution (an example of Markus and Yamabe).
- $\operatorname{Re}\lambda\le0$ is not enough for stability. $A=\begin{pmatrix}0&1\\0&0\end{pmatrix}$ has eigenvalues $0,0$, but $x(t)=(c_1+c_2t,\ c_2)^T$ is unbounded: the eigenvalue $0$ is not semisimple (Theorem 1.27(2)).
- Linearisation decides nothing at a centre. $x'=-y+\mu x(x^2+y^2)$, $y'=x+\mu y(x^2+y^2)$ has a centre as linearisation for every $\mu$, but $r'=\mu r^3$: the origin is an asymptotically stable spiral for $\mu\lt0$ and unstable for $\mu\gt0$.
- Stable is weaker than asymptotically stable. Centres are stable without attracting; so is the origin of $A=\begin{pmatrix}0&0\\0&-1\end{pmatrix}$, which lies on a whole line of equilibria.
- Neither $\tau$ nor $\Delta$ alone decides. $\tau\lt0$ with $\Delta\lt0$ is a saddle ($\operatorname{diag}(1,-3)$); $\Delta\gt0$ with $\tau\gt0$ is unstable. Asymptotic stability needs both $\tau\lt0$ and $\Delta\gt0$.
- Complex pairs: one eigenvector is enough. The real and imaginary parts of $e^{\lambda t}v$ already give two real solutions; using $\bar\lambda$ as well adds nothing new.
Exercises
- 1 Write $y'''-3y''+yy'=e^t$, $y(0)=1$, $y'(0)=0$, $y''(0)=2$ as a first-order system and explain why it has a unique local solution.
- 2 Eliminate $y$ from $x'=x+y$, $y'=4x+y$ to get a second-order equation for $x$, and find the general solution of the system.
- 3 Show that $f(x)=(x_2,\,-\sin x_1)^T$ is Lipschitz on $\mathbb{R}^2$ with constant $1$ in the Euclidean norm, and deduce that every solution of the pendulum equation $y''+\sin y=0$ exists for all $t\in\mathbb{R}$.
- 4 Without solving, find the Wronskian $W(t)$ of the fundamental matrix of $x'=\begin{pmatrix}t&1\\\sin t&-1\end{pmatrix}x$ with $W(0)=2$.
- 5 Can $(t,1)^T$ and $(t^2,2t)^T$ be solutions on $\mathbb{R}$ of $x'=B(t)x$ with $B$ continuous on $\mathbb{R}$? On $(0,\infty)$?
- 6 Show that if solutions $\varphi_1,\varphi_2$ of (H) have $W\equiv0$, then they are linearly dependent, and contrast with Pitfall 1.
- 7 Let $\Phi$ be a fundamental matrix of $x'=A(t)x$. Prove that $\left(\Phi^{-1}\right)^T$ is a fundamental matrix of the adjoint system $z'=-A(t)^Tz$, and that $z^Tx$ is constant whenever $x$ solves (H) and $z$ solves the adjoint system.
- 8 Let $A(t)$ be continuous on $[0,\infty)$ with $\operatorname{tr}A(t)=0$ for all $t$. Prove that $x'=A(t)x$ has a solution that does not tend to $0$ as $t\to\infty$.
- 1 Compute $e^{tA}$ for $A=\begin{pmatrix}0&1\\-4&0\end{pmatrix}$.
- 2 Compute $e^{tA}$ for $A=\begin{pmatrix}4&-1\\1&2\end{pmatrix}$ and solve $x'=Ax$, $x(0)=(1,2)^T$.
- 3 Compute $e^{tA}$ for $A=\begin{pmatrix}1&1\\-1&1\end{pmatrix}$.
- 4 Let $N=\begin{pmatrix}0&1&0\\0&0&1\\0&0&0\end{pmatrix}$. Find $e^{tN}$ and $e^{t(2I+N)}$.
- 5 Let $A=\begin{pmatrix}0&1&1\\1&0&1\\1&1&0\end{pmatrix}$. Find the eigenvalues, $e^{tA}$, and explain why no factor $te^{-t}$ appears although $-1$ is a double eigenvalue.
- 6 Let $A$ be real and skew-symmetric. Prove that $e^{tA}$ is orthogonal and that every solution of $x'=Ax$ has constant Euclidean norm.
- 7 Show that there is no real $2\times2$ matrix $A$ with $e^{A}=\operatorname{diag}(1,-1)$, but that there is a complex one.
- 8 Use the interpolation method ($e^{tA}=\alpha_0(t)I+\alpha_1(t)A$) to compute $e^{tA}$ for $A=\begin{pmatrix}1&3\\3&1\end{pmatrix}$.
- 1 Classify the origin (type and stability) for (a) $\begin{pmatrix}2&-5\\1&-2\end{pmatrix}$; (b) $\begin{pmatrix}-1&4\\-1&-5\end{pmatrix}$; (c) $\begin{pmatrix}3&-2\\4&-1\end{pmatrix}$; (d) $\begin{pmatrix}1&4\\2&-1\end{pmatrix}$.
- 2 For every real $k$ classify the origin of $x'=\begin{pmatrix}-1&k\\1&-1\end{pmatrix}x$.
- 3 Prove directly from the eigenvalues that the origin of a planar system $x'=Ax$ is asymptotically stable if and only if $\operatorname{tr}A\lt0$ and $\det A\gt0$.
- 4 For the predator–prey system $x'=x-xy$, $y'=-y+xy$ ($x,y\gt0$ at the interior point) classify $(0,0)$, show that linearisation is inconclusive at $(1,1)$, and use $H=x-\log x+y-\log y$ to show that the orbits near $(1,1)$ are closed.
- 5 Find all real equilibria of $x'=-x+y^2$, $y'=-2y+x^2$ and classify them.
- 6 Show that the origin of $x'=Ax$ with $A=\begin{pmatrix}0&1&0\\0&0&0\\0&0&-1\end{pmatrix}$ is unstable, although no eigenvalue has positive real part.
- 7 Give a $2\times2$ matrix $A\ne0$ for which the origin is stable but not asymptotically stable and is not a centre.
- 8 Show that for $x'=-y+\mu x(x^2+y^2)$, $y'=x+\mu y(x^2+y^2)$ one has $r'=\mu r^3$, and decide the stability of the origin for each $\mu\in\mathbb{R}$.
- 1 Solve $x'=\begin{pmatrix}0&1\\-1&0\end{pmatrix}x+\begin{pmatrix}0\\t\end{pmatrix}$, $x(0)=0$, by variation of parameters.
- 2 Solve $x'=\begin{pmatrix}1&2\\2&1\end{pmatrix}x+\begin{pmatrix}4e^{-t}\\0\end{pmatrix}$, $x(0)=0$.
- 3 Using the fundamental matrix of Example 2, solve $x'=\begin{pmatrix}0&1\\-2/t^2&2/t\end{pmatrix}x+\begin{pmatrix}0\\t\end{pmatrix}$, $x(1)=0$, on $(0,\infty)$.
- 4 Let all eigenvalues of $A$ have negative real parts and let $b$ be continuous and bounded on $[0,\infty)$. Prove that every solution of $x'=Ax+b(t)$ is bounded on $[0,\infty)$.
- 5 For invertible $A$ and a constant vector $c$, show that $x(t)=x^*+e^{tA}(x_0-x^*)$ with $x^*=-A^{-1}c$ solves $x'=Ax+c$, $x(0)=x_0$. Apply it to $A=\begin{pmatrix}2&-1\\3&-2\end{pmatrix}$, $c=(1,0)^T$, $x_0=0$.
- 6 Show that every solution of $x'=\begin{pmatrix}0&1\\-1&0\end{pmatrix}x+\begin{pmatrix}0\\\cos t\end{pmatrix}$ is unbounded on $[0,\infty)$.
- 7 Let $\Phi$ be a fundamental matrix of $x'=A(t)x$ on $[0,\infty)$ with $\|\Phi(t)\|$ bounded and $\int_0^t\operatorname{tr}A(s)\,ds\ge-C$ for all $t\ge0$. Prove that $\|\Phi(t)^{-1}\|$ is bounded on $[0,\infty)$.
- 8 Let $A$ be constant with no eigenvalue in $\{2\pi ik/T:k\in\mathbb{Z}\}$ and let $b$ be continuous and $T$-periodic. Prove that $x'=Ax+b(t)$ has exactly one $T$-periodic solution.
E1. CSIR Part B GATE MA Let $\Phi$ be a fundamental matrix of $x'=\begin{pmatrix}t&e^{t}\\\sin t&2-t\end{pmatrix}x$ with $\det\Phi(0)=3$. Then $\det\Phi(1)$ equals (A) $3$; (B) $3e$; (C) $3e^{2}$; (D) $3e^{1/2}$.
E2. CSIR Part B For the system $x'=2x-5y$, $y'=2x-4y$, the critical point $(0,0)$ is (A) a saddle point; (B) a stable node; (C) an asymptotically stable spiral point; (D) a centre.
E3. CSIR Part B Let $x(t)$ solve $x'=\begin{pmatrix}-2&1\\0&-2\end{pmatrix}x$, $x(0)=\begin{pmatrix}0\\1\end{pmatrix}$. Then the first component $x_1(1)$ equals (A) $0$; (B) $e^{-2}$; (C) $2e^{-2}$; (D) $e^{-1}$.
E4. CSIR Part C GATE MA Let $A$ be a real $n\times n$ matrix and consider $x'=Ax$. Which of the following statements are true? (One or more options may be correct.) (A) If every eigenvalue of $A$ has negative real part, every solution tends to $0$ as $t\to\infty$. (B) If every solution is bounded on $[0,\infty)$, then every eigenvalue has non-positive real part. (C) If every eigenvalue has non-positive real part, every solution is bounded on $[0,\infty)$. (D) If $A\ne0$ and $A^2=0$, some solution is unbounded on $[0,\infty)$.
E5. CSIR Part C Let $A(t)$ be a continuous $2\times2$ matrix function on $\mathbb{R}$ with $\operatorname{tr}A(t)=2t$, let $\varphi_1,\varphi_2$ be solutions of $x'=A(t)x$ on $\mathbb{R}$, and $W(t)=\det[\varphi_1(t)\ \varphi_2(t)]$. Which of the following are true? (A) If $W(0)=1$, then $W(1)=e$. (B) If $W(1)=0$, then $\varphi_1,\varphi_2$ are linearly dependent on $\mathbb{R}$. (C) $W$ can have exactly one zero on $\mathbb{R}$. (D) If $W(0)=1$, then $W(t)\to0$ as $t\to-\infty$.
E6. CSIR Part C Consider the competing-species system $x'=x(3-x-2y)$, $y'=y(2-x-y)$. Which of the following are true? (A) $(0,0)$ is an unstable node. (B) $(1,1)$ is a saddle point. (C) $(3,0)$ is a stable spiral point. (D) The system has exactly four equilibria in $\mathbb{R}^2$.
Suggested plan — 3 lectures × 60 min.
- From equations to systems; linear structure. Theorem 1.17 with the pendulum, Theorem 1.18 presented as "§1.1 with norms" — prove only part (2) and the linear case of (4) on the board. Example 1 (the energy argument). Theorem 1.19, Liouville's formula with its full proof, Example 2 on the board. Project Interactive 1 (modes 1, 3 and 5) and ask students to predict the area before pressing Play.
- The matrix exponential. Theorem 1.22 (prove (3) by the ODE argument), Theorem 1.23, then Theorem 1.24 with the three $2\times2$ formulas derived from Cayley–Hamilton. Examples 3 and 5 by students in pairs, Example 4 on the board. Finish with pitfall 4 ($e^{A+B}$).
- Forcing, geometry and stability. Theorem 1.25 and Example 6; Definitions 1.14–1.15; Theorem 1.26 with Interactive 2 projected — sweep $a$ across the presets so that the class sees the point cross the parabola and the axes. Theorem 1.27, Theorem 1.28 (proof of (1) via Duhamel and Gronwall), Example 7. Exam-corner E2 and E6 live.
Discussion questions. (1) Why can the Wronskian of two arbitrary vector functions vanish at a single point, but not that of two solutions? What exactly breaks for the Euler system at $t=0$? (2) The eigenvalues of $A(t)$ in pitfall 6 have negative real part for every $t$, yet solutions grow. Where does the "frozen coefficient" reasoning go wrong? (3) In the trace–determinant plane, which types fill open regions and which lie on curves? Why does this make saddles, nodes and spirals "robust" and centres fragile — and what does it say about Theorem 1.28(3)?
Misconceptions to watch for. Applying $e^{A+B}=e^Ae^B$ without commutativity, or $\exp\int A$ for variable $A$; believing that a repeated eigenvalue always produces $te^{\lambda t}$ (Exercise II.5); "$\operatorname{Re}\lambda\le0$ ⇒ stable"; deciding stability from $\tau$ alone or $\Delta$ alone; concluding that a nonlinear equilibrium is a centre because its linearisation is one; using both $\lambda$ and $\bar\lambda$ to produce "four" real solutions of a $2\times2$ system.
Suggested homework. I(1, 4, 5, 7), II(2, 4, 5, 7), III(1, 2, 4, 6), IV(1, 3, 6, 8); Exam corner E1, E4, E5.
Research question. If the interactions in a two-component system are not known and we model them by a random matrix $A$, what is the probability that the equilibrium is a saddle, a node or a spiral — and how likely is it to be stable?
Why it matters. Ecologists studying food webs, systems biologists modelling gene circuits and engineers assessing power grids often know the size of the interactions between components but not their exact values. In 1972 the ecologist Robert May asked whether large, richly connected random ecosystems are stable and found, to the surprise of many, that complexity tends to destroy stability. The $2\times2$ case is the place where the question can be answered completely with the tools of this section.
What to do.
- Let the four entries of $A$ be independent standard normal variables. Show that $P(\det A\lt0)=\tfrac12$ exactly (hint: swapping the two rows preserves the distribution and changes the sign of $\det A$), and that $P(\text{asymptotically stable})=\tfrac14$ (hint: $A\mapsto-A$).
- Simulate $10^6$ matrices (seeded random numbers) and estimate the probabilities of real eigenvalues, saddles, nodes, spirals and stability using only $\tau$ and $\Delta$ (Theorem 1.26). Plot the cloud of points $(\tau,\Delta)$ together with the parabola $\Delta=\tau^2/4$.
- Compare your estimate of $P(\text{real eigenvalues})$ with $1/\sqrt2$, and deduce $P(\text{node})$ and $P(\text{spiral})$.
- Repeat for $3\times3$ and $4\times4$ matrices, now testing stability through the eigenvalues numerically (Theorem 1.27). How does $P(\text{asymptotically stable})$ change with $n$? What if you subtract $sI$ (self-regulation of each species) for $s=0.5,1,2$?
Tools: Python with numpy (numpy.linalg.eigvals) or any spreadsheet with a normal random-number generator; Interactive 2 to check individual matrices.
Hand in: A one-page report with the two exact probabilities and their proofs, a table of simulated frequencies with standard errors, the $(\tau,\Delta)$ scatter plot, and a short paragraph on what changes as $n$ grows.
Going further. Read about the probability that all eigenvalues of an $n\times n$ real Gaussian matrix are real (it is $2^{-n(n-1)/4}$, a result of Edelman, Kostlan and Shub) and about May's stability criterion for large random community matrices; explain the role of the circular law for the eigenvalues of random matrices.
🧪 Self-check quiz
Q1. Let $A(t)$ be a continuous $n\times n$ matrix function on an interval $J$. The set of all solutions of $x'=A(t)x$ on $J$ is
Q2. $\Phi$ is a fundamental matrix of $x'=A(t)x$ with $\operatorname{tr}A(t)\equiv-1$ and $\det\Phi(0)=2$. Then $\det\Phi(3)=$
Q3. For $A=\begin{pmatrix}0&-1\\1&0\end{pmatrix}$, the matrix $e^{tA}$ is
Q4. The origin of $x'=-x+3y$, $y'=-3x-y$ is
Q5. For which matrix is the origin of $x'=Ax$ stable but not asymptotically stable?