6.1 First-Order Equations, Existence–Uniqueness and Linear Equations
Exact, linear and Bernoulli equations; the Lipschitz condition, Picard iterates and the Picard–Lindelöf theorem; linear equations of order $n$, the Wronskian and Abel's formula; constant coefficients, undetermined coefficients, variation of parameters and Cauchy–Euler equations — the core ODE toolkit of GATE MA.
A differential equation is a rule for how a quantity changes; solving it means recovering the quantity itself. Two questions come before any formula. Is there a solution at all, and is it the only one? A physical model whose equation has two different solutions with the same initial state cannot predict anything, and an equation whose solution exists only for a short time describes a system that breaks down. The Picard–Lindelöf theorem answers both questions for $y'=f(x,y)$ under a simple condition on $f$, and its counterexamples show precisely what goes wrong without that condition. What do all the solutions look like? For linear equations the answer is complete: the solutions of a homogeneous linear equation of order $n$ form an $n$-dimensional vector space, a Wronskian decides whether $n$ given solutions span it, and for constant coefficients and for Cauchy–Euler equations the space can be written down from the roots of a polynomial.
GATE MA tests this material in all three formats. Computational MCQs and NATs ask you to solve an exact, linear or Bernoulli equation through a point, to evaluate a Cauchy–Euler or constant-coefficient solution at a given $x$, to compute a Picard iterate, a Wronskian from Abel's formula, or the interval of existence the theorem guarantees. MSQs ask which initial value problems have unique solutions, which pairs of functions can solve a second-order linear equation, and which trial function the method of undetermined coefficients needs. This section develops all of it, with proofs of the central results and a counterexample for each hypothesis.
First-Order Equations You Can Solve Exactly
Let $D\subseteq\mathbb{R}^2$ be open and $f:D\to\mathbb{R}$ continuous. A solution of the first-order equation $y'=f(x,y)$ on an interval $I$ is a differentiable function $\varphi:I\to\mathbb{R}$ with $(x,\varphi(x))\in D$ and $\varphi'(x)=f(x,\varphi(x))$ for every $x\in I$. Given $(x_0,y_0)\in D$, the initial value problem (IVP) is
$$y'=f(x,y),\qquad y(x_0)=y_0 ,$$and a solution of it is a solution on an interval containing $x_0$ with $\varphi(x_0)=y_0$. We also write equations in differential form $M(x,y)\,dx+N(x,y)\,dy=0$; a solution is then a curve along which $M+Ny'=0$ (or $Mx'+N=0$ if $x$ is regarded as a function of $y$).
The simplest equations are separable, $y'=g(x)h(y)$. Where $h(y)\ne0$ we may write $\int\frac{dy}{h(y)}=\int g(x)\,dx+C$; every zero $c$ of $h$ gives in addition the constant solution $y\equiv c$, which this formula never produces. (For $y'=2x(1+y)$: $\log|1+y|=x^2+C$ gives $y=Ae^{x^2}-1$ with $A\ne0$, and $y\equiv-1$ is the solution with $A=0$.) Most of the equations GATE asks about are not separable, but become so — or become derivatives of something — after the right manipulation.
Proof. If $F_x=M$ and $F_y=N$ with $M,N\in C^1$, then $F\in C^2$ and Schwarz's theorem on mixed partials (§1.1) gives $M_y=F_{xy}=F_{yx}=N_x$. Conversely, suppose $M_y=N_x$ on the rectangle and define $F$ by the formula; both integrals run along segments that stay in $D$. By the fundamental theorem of calculus $F_y=N(x,y)$. Differentiating under the integral sign (allowed because $N_x$ is continuous on the compact segment),
$$F_x=M(x,y_0)+\int_{y_0}^{y}N_x(x,t)\,dt=M(x,y_0)+\int_{y_0}^{y}M_y(x,t)\,dt ,$$and the last integral equals $M(x,y)-M(x,y_0)$. So $F_x=M$. ∎
The domain matters. On $\mathbb{R}^2\setminus\{0\}$ the form $\dfrac{-y\,dx+x\,dy}{x^2+y^2}$ satisfies $M_y=N_x$ but has no potential there: its potential would be the polar angle $\theta$, which cannot be defined continuously all the way round the origin. This is the conservative-field phenomenon of §2.1.
Let $M,N$ be $C^1$ on a rectangle $D$.
- If $\dfrac{M_y-N_x}{N}=g(x)$ depends on $x$ alone, then $\mu(x)=e^{\int g(x)\,dx}$ is an integrating factor.
- If $\dfrac{N_x-M_y}{M}=h(y)$ depends on $y$ alone, then $\mu(y)=e^{\int h(y)\,dy}$ is an integrating factor.
- If $M$ and $N$ are homogeneous of the same degree $k$ (homogeneous functions, §1.1) and $xM+yN\ne0$ on $D$, then $\mu=\dfrac{1}{xM+yN}$ is an integrating factor.
Proof. (1) With $\mu=\mu(x)$, exactness of $\mu M\,dx+\mu N\,dy$ means $\mu M_y=\mu'N+\mu N_x$, i.e. $\mu'/\mu=(M_y-N_x)/N=g$, which $\mu=e^{\int g}$ satisfies. (2) is the same with the roles of $x$ and $y$ exchanged. (3) Put $D_0=xM+yN$. A direct computation gives
$$\left(\frac{M}{D_0}\right)_y-\left(\frac{N}{D_0}\right)_x=\frac{N(xM_x+yM_y)-M(xN_x+yN_y)}{D_0^2}.$$By Euler's theorem (§1.1) $xM_x+yM_y=kM$ and $xN_x+yN_y=kN$, so the numerator is $kNM-kMN=0$. ∎
The linear first-order equation is the most important special case: it always has an integrating factor depending on $x$ alone, and its solutions exist on the whole interval where the coefficients are continuous.
Linear. Let $P,Q$ be continuous on an open interval $I$, $x_0\in I$, $y_0\in\mathbb{R}$. The IVP $y'+P(x)y=Q(x)$, $y(x_0)=y_0$ has exactly one solution on $I$, namely
$$y(x)=\frac{1}{\mu(x)}\left(y_0+\int_{x_0}^{x}\mu(t)Q(t)\,dt\right),$$where $\mu(x)=\exp\left(\int_{x_0}^{x}P(t)\,dt\right)$ is the integrating factor. The general solution is $y=\mu^{-1}\left(\int\mu Q\,dx+C\right)$.
Bernoulli. For $n\ne0,1$, the substitution $v=y^{1-n}$ turns $y'+P(x)y=Q(x)y^n$ into the linear equation $v'+(1-n)P(x)v=(1-n)Q(x)$ (on any interval where $y\gt0$, or $y\ne0$ when $y^{1-n}$ is defined for negative $y$). If $n\gt0$, $y\equiv0$ is a further solution that the substitution misses.
Proof. $\mu$ is $C^1$, positive, and $\mu'=P\mu$. If $y$ is differentiable on $I$, then $(\mu y)'=\mu(y'+Py)$, so $y$ solves the equation iff $(\mu y)'=\mu Q$ on $I$, iff $\mu(x)y(x)=\mu(x_0)y(x_0)+\int_{x_0}^{x}\mu Q$ for all $x\in I$. With $\mu(x_0)=1$ and $y(x_0)=y_0$ this is the displayed formula: it shows existence on all of $I$ and uniqueness at once. For Bernoulli, $v=y^{1-n}$ gives $v'=(1-n)y^{-n}y'$; substituting $y'=Qy^n-Py$ gives $v'=(1-n)(Q-Pv)$. ∎
Existence and Uniqueness: the Picard–Lindelöf Theorem
An IVP can have no solution (if $f$ is too wild), several solutions, or a solution that exists only for a short time. Continuity of $f$ is enough for existence; uniqueness needs a little more — that $f$ does not change too fast in $y$.
Proof. For fixed $x$, the mean value theorem applied to $y\mapsto f(x,y)$ on the segment gives $f(x,y_1)-f(x,y_2)=f_y(x,\eta)(y_1-y_2)$ for some $\eta$ between $y_1$ and $y_2$, and $|f_y(x,\eta)|\le L$. For the last claim, a closed rectangle about a point of $D$ is compact, so the continuous $f_y$ is bounded on it. ∎
The converse fails: $f(x,y)=|y|$ is Lipschitz with $L=1$ (by the reverse triangle inequality) but has no $y$-derivative at $y=0$. And $f(x,y)=\sqrt{|y|}$ is not Lipschitz near $y=0$: $\dfrac{|f(x,y)-f(x,0)|}{|y-0|}=|y|^{-1/2}$ is unbounded.
The key observation is that $\varphi$ solves the IVP on $I$ if and only if $\varphi$ is continuous and satisfies the integral equation
$$\varphi(x)=y_0+\int_{x_0}^{x}f\left(t,\varphi(t)\right)dt\qquad(x\in I),$$by the fundamental theorem of calculus in both directions. Picard's idea is to solve this fixed-point equation by iteration.
Let $R=\{(x,y):|x-x_0|\le a,\ |y-y_0|\le b\}$ with $a,b\gt0$, and let $f:R\to\mathbb{R}$ be continuous on $R$ and Lipschitz in $y$ on $R$ with constant $L$. Let $M=\max_R|f|\gt0$ and
$$h=\min\left(a,\ \frac bM\right).$$Then the IVP $y'=f(x,y)$, $y(x_0)=y_0$ has exactly one solution $\varphi$ on $I=[x_0-h,x_0+h]$. The Picard iterates are defined on $I$, their graphs lie in $R$, and they converge to $\varphi$ uniformly on $I$, with $|\varphi(x)-\varphi_n(x)|\le\dfrac ML\dfrac{(Lh)^{n+1}}{(n+1)!}e^{Lh}$ for $L\gt0$.
Proof. Step 1 — the iterates stay in $R$. Suppose $\varphi_n$ is continuous on $I$ with $|\varphi_n-y_0|\le b$ (true for $n=0$). Then $t\mapsto f(t,\varphi_n(t))$ is continuous and bounded by $M$, so $\varphi_{n+1}$ is defined and continuous, and $|\varphi_{n+1}(x)-y_0|\le M|x-x_0|\le Mh\le b$. This is where $h\le b/M$ is used: the graph cannot leave $R$ through its top or bottom.
Step 2 — the differences shrink factorially. We show $|\varphi_{k+1}(x)-\varphi_k(x)|\le\dfrac{ML^k|x-x_0|^{k+1}}{(k+1)!}$ on $I$. For $k=0$ it is Step 1. If it holds for $k-1$, then by the Lipschitz condition (take $x\ge x_0$; the other side is symmetric)
$$|\varphi_{k+1}(x)-\varphi_k(x)|\le L\int_{x_0}^{x}|\varphi_k-\varphi_{k-1}|\,dt\le L\int_{x_0}^{x}\frac{ML^{k-1}(t-x_0)^{k}}{k!}\,dt ,$$which is the claim.
Step 3 — uniform convergence to a solution. On $I$ the terms of $\varphi_0+\sum_k(\varphi_{k+1}-\varphi_k)$ are bounded by $\frac ML\frac{(Lh)^{k+1}}{(k+1)!}$, the terms of a convergent series. By the Weierstrass M-test (§4.2) $\varphi_n\to\varphi$ uniformly on $I$; $\varphi$ is continuous with graph in $R$, and the tail of the series gives the error bound. Since $f(t,\varphi_n(t))$ and $f(t,\varphi(t))$ differ by at most $L|\varphi_n(t)-\varphi(t)|$, which tends to $0$ uniformly, we may pass to the limit under the integral in Definition 6.4: $\varphi$ satisfies the integral equation, hence solves the IVP.
Step 4 — uniqueness. Let $\psi$ be another solution on $J\subseteq I$, $x_0\in J$. As in Step 1, its graph stays in $R$: while it is in $R$, $|\psi'|\le M$, so if it left $R$ at some $x_1$ with $|x_1-x_0|\lt h$, at the first such point $b=|\psi(x_1)-y_0|\le M|x_1-x_0|\lt Mh\le b$ — impossible. Put $K=\max_J|\varphi-\psi|$. Then $|\varphi(x)-\psi(x)|\le L\left|\int_{x_0}^{x}|\varphi-\psi|\,dt\right|$, and inserting this bound into itself $n$ times gives $|\varphi(x)-\psi(x)|\le K\dfrac{(L|x-x_0|)^n}{n!}\to0$. So $\psi=\varphi$ on $J$. ∎
The same argument, read as "the map $\varphi\mapsto y_0+\int f(t,\varphi)$ has a unique fixed point", is the contraction mapping principle of §4.1. Two companions complete the picture.
- Peano. If $f$ is merely continuous on $R$ (notation of Theorem 6.5), the IVP has at least one solution on $[x_0-h,x_0+h]$.
- Strip. If $f$ is continuous on the strip $S=[\alpha,\beta]\times\mathbb{R}$ and Lipschitz in $y$ on all of $S$, then for every $x_0\in[\alpha,\beta]$ and $y_0\in\mathbb{R}$ the IVP has exactly one solution, and it is defined on the whole interval $[\alpha,\beta]$. In particular the linear equation $y'=P(x)y+Q(x)$ with $P,Q$ continuous on $[\alpha,\beta]$ has this property ($L=\max|P|$).
Proof (sketch). (1) Euler polygons with step $1/k$ stay in $R$, are uniformly bounded and uniformly Lipschitz (slope at most $M$), so by the Ascoli–Arzelà theorem (§4.2) a subsequence converges uniformly; the limit satisfies the integral equation. (2) On a strip the iterates never need to stay in a box, so Step 1 is unnecessary; Steps 2–4 run on all of $[\alpha,\beta]$ with $M$ replaced by $M_0=\max_{[\alpha,\beta]}|f(x,y_0)|$, since $|\varphi_1-y_0|\le M_0|x-x_0|$. ∎
Linear Equations of Order $n$: the Wronskian and Abel's Formula
Proof (idea). With $Y=(y,y',\dots,y^{(n-1)})^{\mathsf T}$ the equation becomes the first-order system $Y'=A(x)Y+F(x)$, where $A$ is the companion matrix of the $p_k$. On a compact subinterval $[\alpha,\beta]\subset J$ the right side is Lipschitz in $Y$ with constant $\max\|A(x)\|$ on the whole strip $[\alpha,\beta]\times\mathbb{R}^n$, and the vector form of Theorem 6.6(2) applies; exhausting $J$ by such subintervals gives a solution on $J$. ∎
Proof. (a) $L$ is linear, so $V=\ker L$ is a vector space. Fix $x_0$ and define $\Phi:V\to\mathbb{R}^n$, $\Phi(y)=(y(x_0),\dots,y^{(n-1)}(x_0))$. $\Phi$ is linear; it is onto by the existence part of Theorem 6.7 and one-to-one by the uniqueness part. Hence $\dim V=n$. (b) $L[y-y_p]=f-f=0$. ∎
Let $y_1,\dots,y_n$ be solutions of $L[y]=0$ on $J$ and $W=W(y_1,\dots,y_n)$. Then for any $x_0\in J$
$$W(x)=W(x_0)\exp\left(-\int_{x_0}^{x}p_{n-1}(t)\,dt\right)\qquad(\text{Abel's formula}).$$Consequently $W$ is either identically zero or never zero on $J$, and $y_1,\dots,y_n$ form a fundamental set $\iff W(x_0)\ne0$ for some $x_0\in J$ $\iff W(x)\ne0$ for every $x\in J$.
Proof. Abel, $n=2$. $W'=(y_1y_2'-y_1'y_2)'=y_1y_2''-y_1''y_2$. Substituting $y_k''=-p_1y_k'-p_0y_k$, the $p_0$ terms cancel and $W'=-p_1(y_1y_2'-y_1'y_2)=-p_1W$. This linear first-order equation has the solution displayed (Theorem 6.3). General $n$. Differentiating the determinant row by row, every term but the last has two equal rows, so $W'$ is the determinant with last row $(y_1^{(n)},\dots,y_n^{(n)})$. Replace $y_k^{(n)}$ by $-\sum_jp_jy_k^{(j)}$ and add to the last row $p_j$ times the row of $j$-th derivatives, for $j\le n-2$: what remains in the last row is $-p_{n-1}$ times the row of $(n-1)$-th derivatives, so $W'=-p_{n-1}W$.
Test. If $W(x_0)=0$, the columns of the Wronski matrix at $x_0$ are dependent: some $c\ne0$ has $\sum c_ky_k^{(j)}(x_0)=0$ for $0\le j\le n-1$. Then $y=\sum c_ky_k$ solves $L[y]=0$ with zero initial data at $x_0$, so $y\equiv0$ by Theorem 6.7 — the $y_k$ are dependent. Conversely, if $\sum c_ky_k\equiv0$ with $c\ne0$, differentiating $n-1$ times shows the Wronski matrix kills $c$ at every point, so $W\equiv0$. The dichotomy is visible in Abel's formula, since the exponential never vanishes. ∎
For arbitrary functions (not solutions of one normal equation) "$W\equiv0$" does not imply dependence; Example 5(b) shows this. When one solution of a second-order equation is known, Abel's formula hands us a second.
Proof. Let $y_2$ be any solution independent of $y_1$. By Abel, $W(y_1,y_2)=Ce^{-\int p}$ with $C\ne0$, and $\left(\dfrac{y_2}{y_1}\right)'=\dfrac{W}{y_1^2}$. Integrating, $y_2/y_1$ is $C\int e^{-\int p}/y_1^2$ plus a constant; dropping the constants (which only add multiples of $y_1$) gives the formula. Directly: the displayed $y_2$ has $W(y_1,y_2)=y_1^2(y_2/y_1)'=e^{-\int p}\ne0$, and one checks that it solves the equation. ∎
Constant Coefficients, Undetermined Coefficients, Variation of Parameters
Write $D=\dfrac{d}{dx}$. A homogeneous equation with constant real coefficients is $p(D)y=0$, where $p(r)=r^n+a_{n-1}r^{n-1}+\dots+a_0$. Everything rests on the shift rule $(D-r)\left(e^{rx}u\right)=e^{rx}u'$ — a one-line product-rule computation — and on the fact that polynomials in $D$ with constant coefficients multiply like ordinary polynomials.
Proof. They are solutions. Write $p(D)=q(D)(D-r_j)^{m_j}$. By the shift rule applied $m_j$ times, $(D-r_j)^{m_j}\left(x^ie^{r_jx}\right)=e^{r_jx}D^{m_j}x^i=0$ for $i\lt m_j$; hence $p(D)\left(x^ie^{r_jx}\right)=0$. Independence (sketch). Suppose $\sum_jP_j(x)e^{r_jx}\equiv0$ with polynomials $P_j$ of degree $\lt m_j$, not all zero; we may assume $P_k\ne0$. Divide by $e^{r_1x}$ and differentiate $\deg P_1+1$ times: the $P_1$ term disappears, and each other term becomes $Q_j(x)e^{(r_j-r_1)x}$ with $\deg Q_j=\deg P_j$ (because $r_j\ne r_1$). Induction on the number of distinct roots gives a contradiction. There are $n$ functions, so by Theorem 6.8 they span. Real form. $e^{(\alpha\pm i\beta)x}=e^{\alpha x}(\cos\beta x\pm i\sin\beta x)$, and half-sums and half-differences of solutions are solutions. ∎
For a forcing term that is itself a solution of some constant-coefficient equation — a polynomial times an exponential times a sine or cosine — a particular solution can be guessed in the right form.
Proof (sketch). $f$ is annihilated by $q(D)=\left((D-a)^2+b^2\right)^{d+1}$ (by $(D-a)^{d+1}$ if $b=0$), by Theorem 6.11. Any solution of $p(D)y=f$ therefore solves $q(D)p(D)y=0$, whose fundamental set (Theorem 6.11) consists of the solutions of $p(D)y=0$ together with $x^{s+i}e^{ax}\cos bx$, $x^{s+i}e^{ax}\sin bx$, $0\le i\le d$ — the root $a+ib$ now has multiplicity $s+d+1$. Discarding the solutions of $p(D)y=0$ leaves the stated form, and the coefficients are found by substitution. ∎
Let $y_1,y_2$ be a fundamental set of $y''+p(x)y'+q(x)y=0$ on $J$, with $W=W(y_1,y_2)$, and let $f$ be continuous on $J$. Then a particular solution of $y''+p(x)y'+q(x)y=f(x)$ — in normal form — is $y_p=u_1y_1+u_2y_2$ with
$$u_1'=-\frac{y_2f}{W},\qquad u_2'=\frac{y_1f}{W}.$$For order $n$: $y_p=\sum u_ky_k$ with $u_k'=\dfrac{W_k}{W}f$, where $W_k$ is $W$ with its $k$-th column replaced by $(0,\dots,0,1)^{\mathsf T}$.
Proof. Look for $y_p=u_1y_1+u_2y_2$ and impose $u_1'y_1+u_2'y_2=0$. Then $y_p'=u_1y_1'+u_2y_2'$ and $y_p''=u_1y_1''+u_2y_2''+u_1'y_1'+u_2'y_2'$. Substituting, and using that $y_1,y_2$ solve the homogeneous equation, $y_p''+py_p'+qy_p=u_1'y_1'+u_2'y_2'$. So $y_p$ works if
$$\begin{pmatrix}y_1&y_2\\y_1'&y_2'\end{pmatrix}\begin{pmatrix}u_1'\\u_2'\end{pmatrix}=\begin{pmatrix}0\\f\end{pmatrix},$$a system with determinant $W\ne0$; Cramer's rule gives the formulas, and $u_1,u_2$ exist because the right-hand sides are continuous. The order-$n$ case is the same with the conditions $\sum u_k'y_k^{(j)}=0$ for $j\le n-2$. ∎
- Homogeneous part. Constant coefficients: factor $p(r)$ and write the fundamental set (Theorem 6.11). Cauchy–Euler: use $x=e^t$ (next subsection). Otherwise look for one solution and reduce the order (Theorem 6.10).
- Particular solution. If $f$ is of the form $e^{ax}(P\cos bx+Q\sin bx)$, use undetermined coefficients, multiplying by $x^s$ when $a+ib$ is a root of multiplicity $s$; split sums of such terms. For anything else ($\sec x$, $\log x$, $\frac{e^x}{1+x^2}$, …) use variation of parameters, after putting the equation in normal form.
- Initial conditions last. Apply them to $y_c+y_p$, never to $y_c$ alone.
- Check by substitution, and with Abel's formula when a Wronskian is involved.
Cauchy–Euler Equations
When the coefficient of $y^{(k)}$ is a constant multiple of $x^k$, the equation does not change if $x$ is replaced by $\lambda x$; such equidimensional equations become constant-coefficient equations in the variable $t=\log x$.
On $x\gt0$ put $x=e^t$, $Y(t)=y(e^t)$, and $\theta=\dfrac{d}{dt}$. Then for every $k\ge1$
$$x^k\frac{d^ky}{dx^k}=\theta(\theta-1)\cdots(\theta-k+1)Y,$$so the Cauchy–Euler equation becomes a constant-coefficient equation for $Y$ with forcing $f(e^t)$. For $n=2$, with roots $r_1,r_2$ of the indicial equation, a fundamental set on $x\gt0$ is
- $x^{r_1},\ x^{r_2}$ if $r_1\ne r_2$ are real;
- $x^{r},\ x^{r}\log x$ if $r_1=r_2=r$;
- $x^{\alpha}\cos(\beta\log x),\ x^{\alpha}\sin(\beta\log x)$ if $r=\alpha\pm i\beta$, $\beta\ne0$.
On $x\lt0$ the same holds with $x$ replaced by $|x|$. More generally $(px+q)^ny^{(n)}+\dots=f$ (a Legendre linear equation) reduces the same way with $px+q=e^t$.
Proof. By the chain rule $\dfrac{dY}{dt}=y'(e^t)e^t$, i.e. $xy'=\theta Y$. Suppose $x^ky^{(k)}=\theta(\theta-1)\cdots(\theta-k+1)Y$. Differentiating $x^ky^{(k)}$ with respect to $x$ gives $kx^{k-1}y^{(k)}+x^ky^{(k+1)}$, so
$$x^{k+1}y^{(k+1)}=x\frac{d}{dx}\left(x^ky^{(k)}\right)-kx^ky^{(k)}=(\theta-k)\left(x^ky^{(k)}\right),$$because $x\frac{d}{dx}=\theta$. This is the case $k+1$. For $n=2$ the transformed equation is $\theta(\theta-1)Y+a\theta Y+bY=f(e^t)$, whose characteristic polynomial is the indicial polynomial; Theorem 6.11 in the variable $t=\log x$ gives the list. On $x\lt0$ put $x=-e^t$; the same computation goes through. ∎
- Continuity gives existence, not uniqueness. $y'=3y^{2/3}$, $y(0)=0$ has infinitely many solutions; so does $y'=2\sqrt{|y|}$, $y(0)=0$. Neither $f$ is Lipschitz near $y=0$.
- Lipschitz is sufficient, not necessary. $y'=y\log|y|$ (with $f=0$ at $y=0$) is not Lipschitz near $y=0$, yet $y(0)=0$ has only the zero solution: the nonzero solutions $\pm\exp(Ce^x)$ never reach $0$.
- Lipschitz $\not\Rightarrow$ differentiable in $y$. $f=|y|$ is Lipschitz with $L=1$; the IVP $y'=|y|$, $y(0)=0$ has the unique solution $y\equiv0$.
- Local means local. $y'=y^2$, $y(0)=1$: $f$ is a polynomial, but the solution $\frac1{1-x}$ dies at $x=1$. Theorem 6.5 guarantees at most $|x|\le\tfrac14$ here; the true interval is $(-\infty,1)$.
- The exactness test needs a simply connected domain. $\frac{-y\,dx+x\,dy}{x^2+y^2}$ passes the test on $\mathbb{R}^2\setminus\{0\}$ but has no potential there.
- Dividing loses solutions. Separating $y'=y^2$ or substituting $v=y^{1-n}$ in a Bernoulli equation misses $y\equiv0$; an integrating factor that vanishes or blows up somewhere can add or remove curves.
- $W\equiv0$ does not imply dependence for arbitrary functions: $x^2$ and $x|x|$ on $(-1,1)$. It does for solutions of one normal linear equation (Theorem 6.9).
- Solutions' Wronskians cannot vanish at a single point. $x$ and $\sin x$ (Wronskian zero at $0$, not at $\pi$) cannot both solve a normal second-order linear equation on $\mathbb{R}$; nor can two solutions with a common zero, or a common critical point, be independent.
- Abel and variation of parameters need the normal form. For $x^2y''-2xy'+2y=x^3$ use $p=-\frac2x$ and $f=x$.
- Resonance. For $y''+4y=\sin2x$ the trial $A\cos2x+B\sin2x$ fails (it solves the homogeneous equation); multiply by $x$.
- Superposition is for linear equations only. $0$ and $\frac{1}{1-x}$ solve $y'=y^2$; their sum $\frac1{1-x}$ does, but $2\cdot\frac1{1-x}$ does not.
- Cauchy–Euler lives on $x\gt0$ (or $x\lt0$). The equation is singular at $0$; $x^r$ must be read as $|x|^r$ on $x\lt0$, and $\log x$ as $\log|x|$.
- CSIR Unit 3 · §1.1 Existence and Uniqueness of Solutions — the integral equation, the complete Picard–Lindelöf proof with its error estimate, Peano's theorem, maximal solutions and the blow-up alternative, Gronwall's inequality and continuous dependence on the data.
- CSIR Unit 3 · §2.1 General Theory of Linear Equations — existence for linear systems, the dimension theorem, the Wronskian for order $n$ and Abel's formula in full, independence of the constant-coefficient fundamental set, Euler–Cauchy equations and the Sturm separation theorem.
- CSIR Unit 3 · §2.2 Non-Homogeneous Equations and Variation of Parameters — annihilators and the proof of the method of undetermined coefficients, variation of parameters for order $n$, and the operator method $1/f(D)$.
Two Labs
Exercises
- 1 Solve $(e^x\sin y-2x)\,dx+{}$ $(e^x\cos y+1)\,dy=0$, $y(0)=0$.
- 2 Find an integrating factor of $y\,dx+(2x-ye^y)\,dy=0$ ($y\gt0$) and solve the equation.
- 3 Use Theorem 6.2(3) — or the substitution $y=vx$ — to solve $(x^2+y^2)\,dx-2xy\,dy=0$ on $x\gt0$.
- 4 Solve $y'+y\tan x=\sec x$ on $\left(-\tfrac\pi2,\tfrac\pi2\right)$ with $y(0)=1$.
- 5 Solve the Bernoulli equation $y'+y=e^xy^2$.
- 6 Modelling. A 100-litre tank initially holds pure water. Brine with $0.5$ kg of salt per litre flows in at $4$ L/min and the well-mixed solution flows out at the same rate. Find the salt content $A(t)$ and the time at which it reaches $25$ kg.
- 7 Solve $y'=(x+y)^2$ with the substitution $v=x+y$.
- 8 Let $\mu$ be an integrating factor of $M\,dx+N\,dy=0$ and $F$ a potential of $\mu M\,dx+\mu N\,dy$. Prove that $\mu\,g(F)$ is again an integrating factor for every continuous nowhere-zero $g$. (So integrating factors are never unique.)
- 1 Show that $f(x,y)=x^2|y|$ is Lipschitz in $y$ on $[-1,1]\times\mathbb{R}$, although $\partial f/\partial y$ does not exist on the line $y=0$.
- 2 Find a Lipschitz constant for $f(x,y)=\sin(xy)$ on $[0,2]\times\mathbb{R}$.
- 3 Compute the Picard iterates $\varphi_1,\varphi_2,\varphi_3$ for $y'=x+y$, $y(0)=1$, and compare with the Taylor series of the exact solution.
- 4 For $y'=x^2+y^2$, $y(0)=0$ and rectangles $|x|\le a$, $|y|\le b$, find the largest $h$ that Theorem 6.5 can guarantee.
- 5 How many solutions does $y'=2\sqrt{|y|}$, $y(0)=0$ have on $\mathbb{R}$? Describe them all.
- 6 Show that $y'=x+\sin y$, $y(0)=0$ has a unique solution defined on all of $\mathbb{R}$.
- 7 Solve $y'=1+y^2$, $y(0)=0$, find its maximal interval of existence, and compare with the best $h$ of Theorem 6.5.
- 8 Let $f$ and $f_y$ be continuous on $\mathbb{R}^2$. Prove that the graphs of two different solutions of $y'=f(x,y)$ cannot meet. Deduce that the solution of $y'=y(1-y)$, $y(0)=\tfrac12$, satisfies $0\lt y(x)\lt1$ wherever it is defined.
- 1 Compute $W\left(e^x,e^{2x},e^{3x}\right)$.
- 2 $y_1,y_2$ solve $(1-x^2)y''-2xy'+6y=0$ on $(-1,1)$ with $W(y_1,y_2)(0)=2$. Find $W(x)$ and $W\left(\tfrac12\right)$.
- 3 $y_1=x$ solves $x^2y''-x(x+2)y'+(x+2)y=0$ on $x\gt0$. Find a second solution by reduction of order.
- 4 Solve $y'''-y''-y'+y=0$.
- 5 Solve $y^{(4)}+8y''+16y=0$.
- 6 Find the general solution of $y''+2y'+5y=10\cos x$.
- 7 Use variation of parameters to find a particular solution of $y''-y=\dfrac{2}{1+e^x}$.
- 8 Modelling. A $1$ kg mass on a spring of stiffness $9$ N/m moves with damping force $-c\,x'$. Find the critical damping coefficient, and the motion when $c$ takes this value, $x(0)=0.1$ m and $x'(0)=0$. Does the mass ever pass through equilibrium?
- 9 Let $y_1,y_2$ solve $y''+p(x)y'+q(x)y=0$ on $J$ ($p,q$ continuous). Prove: if both have a local maximum at the same point $x_0\in J$, they are linearly dependent.
- 1 Solve $x^2y''+xy'-9y=0$.
- 2 Solve $x^2y''-xy'+5y=0$.
- 3 Solve $x^3y'''+3x^2y''-2xy'+2y=0$.
- 4 Find the general solution of $x^2y''-2y=3x^2$.
- 5 Solve $(x+1)^2y''-3(x+1)y'+4y=0$ on $x\gt-1$.
- 6 Solve $x^2y''+3xy'+y=0$, $y(1)=1$, $y'(1)=0$, and find $y(e)$.
- 7 Prove that every nontrivial solution of $x^2y''+xy'+y=0$ is bounded on $(0,\infty)$, has infinitely many zeros in $(0,1)$, and has only finitely many zeros in $[1,10]$.
E1. GATE MA · MCQ Let $y(x)$ be the solution of $\left((2x+3)y+e^{-2x}\right)dx+(x+1)\,dy=0$ with $y(0)=1$, on an interval containing $[0,1]$. Then $y(1)$ equals (A) $0$; (B) $\dfrac{1}{2e^2}$; (C) $e^{-2}$; (D) $\tfrac12$.
E2. GATE MA · MCQ For the IVP $y'=e^{-x^2}+y^2$, $y(0)=0$, on the rectangle $R=\{|x|\le2,\ |y|\le3\}$, the Picard–Lindelöf theorem guarantees a unique solution on $|x|\le h$ with $h$ equal to (A) $0.3$; (B) $\tfrac13$; (C) $2$; (D) $0.2$.
E3. GATE MA · MSQ CSIR Part C Which of the following statements is/are true? (One or more options may be correct.) (A) If $y_1,y_2$ solve $y''+p(x)y'+q(x)y=0$ on $(0,\infty)$ with $p,q$ continuous, and $W(y_1,y_2)(1)=0$, then $W(y_1,y_2)\equiv0$ on $(0,\infty)$. (B) There are continuous $p,q$ on $(0,\infty)$ such that $x$ and $x^2$ both solve $y''+p(x)y'+q(x)y=0$ there. (C) There are continuous $p,q$ on $(0,\infty)$ such that $x$ and $\sin x$ both solve $y''+p(x)y'+q(x)y=0$ there. (D) Two solutions of $y''+p(x)y'+q(x)y=0$ ($p,q$ continuous on $J$) with a common zero in $J$ are linearly dependent.
E4. GATE MA · MSQ Which of the following initial value problems has/have a unique solution defined on all of $\mathbb{R}$? (One or more options may be correct.) (A) $y'=|y|$, $y(0)=0$. (B) $y'=\sqrt{|y|}$, $y(0)=0$. (C) $y'=y^2$, $y(0)=1$. (D) $y'=\sin y+x^2$, $y(0)=5$.
E5. GATE MA · NAT Let $y$ solve $x^2y''+xy'-4y=4x^2$ on $x\gt0$ with $y(1)=1$, $y'(1)=0$. The value of $y(2)$ is ______ (correct to two decimal places).
E6. GATE MA · NAT For the IVP $y'=x+y^2$, $y(0)=1$, let $\varphi_2$ be the second Picard iterate (with $\varphi_0\equiv1$). The value of $\varphi_2(1)$ is ______ (correct to two decimal places).
Suggested plan — 3 lectures × 60 min.
- Solving first-order equations. Definitions 6.1–6.2, Theorem 6.1 with proof, Theorem 6.2 (prove part 1 on the board, leave part 3 as reading), Theorem 6.3, Examples 1 and 2. End with the question "why did $y$ in Example 2(b) die at $x=2$ when the linear equation for $v$ did not?" — it motivates the next lecture.
- Existence and uniqueness. The integral equation, Definition 6.4, Theorem 6.5 with Steps 1–4 (Step 2 is the heart; do it carefully). Project Interactive 1: first $y'=y$ (iterates are Taylor polynomials), then $y'=1+y^2$ (convergence beyond the green band, failure near $\pm\tfrac\pi2$), then $y'=3y^{2/3}$ (the iterates pick the zero solution and never notice $x^3$). Examples 3 and 4; GATE corner E2 and E4 live.
- Linear equations. Theorems 6.7–6.9 (prove Abel for $n=2$), Example 5, Theorem 6.11 via the shift rule, Theorems 6.12–6.13, Example 6, Theorem 6.14 and Example 7. Project Interactive 2 in both modes and ask the class to predict the solution shape from the roots before releasing the slider.
Discussion questions. (1) Picard iterates for $y'=3y^{2/3}$, $y(0)=0$ converge — to which solution, and why does that not contradict Example 4(a)? (2) Why can two solutions of $y''+py'+qy=0$ never have a common critical point unless they are proportional, while $x^2$ and $x^3$ (not solutions of one normal equation on $\mathbb{R}$) do? (3) In Interactive 2, the same roots give $e^{rx}$ in one mode and $x^r$ in the other. Why are the zeros of the oscillating Cauchy–Euler solution spaced further and further apart?
Misconceptions to watch for. Applying the exactness test on a punctured plane; dividing by $y$ and losing $y\equiv0$; confusing "Lipschitz" with "differentiable"; reading Theorem 6.5's $h$ as the true interval; using Abel's formula or variation of parameters without the normal form; forgetting the factor $x^s$ at resonance; imposing initial conditions on $y_c$ alone; writing $x^r$ on $x\lt0$.
Suggested homework. I(2, 3, 5, 6, 8), II(3, 4, 5, 7, 8), III(2, 3, 6, 7, 9), IV(3, 4, 6, 7); GATE corner E1, E3, E5.
Research question. A tank drains through a small hole in its base. Looking at the empty tank, can you tell when it became empty — and how should a vessel be shaped so that its water level falls at a constant rate, as in the water clocks of ancient Egypt, Greece and China?
Why it matters. Torricelli's law says that water leaves a hole of area $a$ with speed $\sqrt{2gh}$, where $h$ is the depth. For a tank of cross-section $A(h)$ this gives $A(h)\,h'=-a\sqrt{2gh}$, an equation whose right-hand side is not Lipschitz at $h=0$. Forward in time the solution is unique; backward in time it is not. The same loss of uniqueness appears in models of friction, of fluid draining from porous rock, and in the "finite-time stopping" controllers used in robotics.
What to do.
- For a cylinder ($A$ constant) solve $h'=-k\sqrt h$, $h(0)=H$, and show that the tank is empty at $T=2\sqrt H/k$. Write down all solutions on $\mathbb{R}$ with $h(t_1)=0$ for a given $t_1$, and explain why they show non-uniqueness backward in time.
- Check which hypothesis of Theorem 6.5 fails at $h=0$, and why it does not fail at any $h\gt0$.
- Find the profile $r(h)$ of a vessel of revolution for which $h'$ is constant. (You should get $r\propto h^{1/4}$.) Sketch it and compare with pictures of historical water clocks.
- Run Euler's method for $h'=-\sqrt{|h|}$ backward in time from $h=0$ with several step sizes. What does the numerical method "choose", and why?
Tools: pencil and paper for 1–3; Python or a spreadsheet for 4.
Hand in: the solution family of step 1 with a sketch, the vessel profile of step 3, a table or plot from step 4, and a paragraph answering the research question.
Going further. Read about Osgood's uniqueness criterion: $y'=f(y)$ has unique solutions through a zero $y_0$ of $f$ if $\int dy/|f(y)|$ diverges near $y_0$. Check it on $\sqrt{|y|}$, $|y|$ and $y\log|y|$.
🧪 Self-check quiz
Q1. An integrating factor of $xy'+3y=x^2$ on $x\gt0$ (written in the form $y'+\frac3xy=x$) is
Q2. $y_1,y_2$ solve $y''+2xy'+(\sin x)y=0$ on $\mathbb{R}$ and $W(y_1,y_2)(0)=3$. Then $W(y_1,y_2)(1)$ equals
Q3. The general solution of $x^2y''-3xy'+4y=0$ on $x\gt0$ is
Q4. The method of undetermined coefficients for $y''-4y'+4y=e^{2x}$ needs the trial function
Q5. How many solutions does the IVP $y'=3y^{2/3}$, $y(0)=0$ have on $\mathbb{R}$?