Chapter 1 · First-Order Ordinary Differential Equations

1.1 Existence and Uniqueness of Solutions

When does $y'=f(x,y)$, $y(x_0)=y_0$ have a solution, when is it the only one, and how long does it live? Lipschitz conditions, Picard iterates, Peano's theorem, blow-up and Gronwall's inequality.

A differential equation $y'=f(x,y)$ together with a starting value $y(x_0)=y_0$ is the mathematical form of a deterministic law: the present state and the rule of change should fix the whole future. Whether they really do is not obvious. The equation $y'=3y^{2/3}$ with $y(0)=0$ is satisfied by $y\equiv0$ and by $y=x^3$, and by infinitely many other functions; the perfectly smooth equation $y'=y^2$ with $y(0)=1$ has a solution that runs off to infinity at $x=1$; and a discontinuous right-hand side can leave an initial value problem with no solution at all. Before solving equations we must know which questions have answers.

This section answers them. The Lipschitz condition gives existence and uniqueness (the Picard–Lindelöf theorem), proved constructively by Picard's successive approximations — the same iteration that reappears in Chapter 5 as a numerical method and in Chapter 7 for Volterra integral equations. Continuity alone gives existence without uniqueness (Peano). Solutions extend until they blow up or leave the domain, and Gronwall's inequality shows that they depend continuously on the data. CSIR NET examines exactly these points: the interval $h=\min(a,b/M)$, whether a given $f$ is Lipschitz, how many solutions an IVP has, and the maximal interval of existence. At the end we collect the classical solution methods — exact equations, integrating factors, linear and Bernoulli equations — as tools for computing the solutions the theory promises.

Initial Value Problems and the Integral Equation

📘 Definition 1.1 — Initial value problem and solution

Let $D\subseteq\mathbb{R}^2$ and let $f:D\to\mathbb{R}$ be continuous. The initial value problem (IVP) with data $(x_0,y_0)\in D$ is

$$y'=f(x,y),\qquad y(x_0)=y_0.$$

A solution on an interval $I\ni x_0$ is a differentiable function $\varphi:I\to\mathbb{R}$ whose graph lies in $D$, i.e. $(x,\varphi(x))\in D$ for all $x\in I$, with $\varphi'(x)=f(x,\varphi(x))$ for every $x\in I$ and $\varphi(x_0)=y_0$. Geometrically, the graph of $\varphi$ is an integral curve: at each of its points its slope is the value of $f$ there.

Since $\varphi'(x)=f(x,\varphi(x))$ is a composite of continuous functions, every solution is automatically $C^1$. The first step of the whole theory is to trade the derivative for an integral, because integrals behave far better under limits.

📐 Theorem 1.1 — Equivalence with an integral equation
Let $f$ be continuous on $D$ and let $\varphi:I\to\mathbb{R}$ be continuous with graph in $D$. Then $\varphi$ is a solution of the IVP on $I$ if and only if $$\varphi(x)=y_0+\int_{x_0}^{x}f\big(t,\varphi(t)\big)\,dt\qquad\text{for all }x\in I.$$

Proof. If $\varphi$ solves the IVP, then $\varphi'$ is continuous, and the fundamental theorem of calculus gives $\varphi(x)-\varphi(x_0)=\int_{x_0}^x\varphi'(t)\,dt=\int_{x_0}^x f(t,\varphi(t))\,dt$, with $\varphi(x_0)=y_0$. Conversely, if $\varphi$ is continuous and satisfies the integral equation, the integrand $t\mapsto f(t,\varphi(t))$ is continuous, so the right-hand side is differentiable with derivative $f(x,\varphi(x))$ (fundamental theorem of calculus); hence $\varphi$ is differentiable, $\varphi'(x)=f(x,\varphi(x))$, and putting $x=x_0$ gives $\varphi(x_0)=y_0$. ∎

Notice the gain: the integral equation asks only for continuity of $\varphi$, and differentiability comes for free. A limit of a uniformly convergent sequence of continuous functions is continuous, but need not be differentiable — so it is the integral form that lets us build solutions as limits.

✏️ Example 1 — From an integral equation to an IVP
Find all continuous functions $y$ on $\mathbb{R}$ with $\displaystyle y(x)=2+\int_1^x t\,y(t)\,dt$.
Convert By Theorem 1.1 (with $f(x,y)=xy$, continuous on $\mathbb{R}^2$), a continuous $y$ satisfies the equation iff it solves $y'=xy$, $y(1)=2$. Putting $x=1$ in the integral equation gives the initial condition; differentiating gives the equation.
Solve The equation is linear: $\left(y\,e^{-x^2/2}\right)'=e^{-x^2/2}(y'-xy)=0$, so $y=Ke^{x^2/2}$, and $y(1)=2$ gives $K=2e^{-1/2}$: $$y(x)=2e^{(x^2-1)/2}.$$
Uniqueness The step "$\left(ye^{-x^2/2}\right)'=0\Rightarrow$ constant" shows that this is the only continuous solution — no general theorem needed here. Theorem 1.6 below gives the same conclusion for every linear equation.
Check $\int_1^x t\cdot2e^{(t^2-1)/2}dt=\left[2e^{(t^2-1)/2}\right]_1^x=2e^{(x^2-1)/2}-2$, and adding $2$ returns $y(x)$.

The Lipschitz Condition

Continuity of $f$ will turn out to be enough for existence. For uniqueness we need control on how fast $f$ changes in the $y$-direction: two solutions that start together separate at a rate governed by $f(x,y_1)-f(x,y_2)$, and the Lipschitz condition bounds this by a multiple of $|y_1-y_2|$.

📘 Definition 1.2 — Lipschitz condition in $y$

A function $f:D\to\mathbb{R}$ satisfies a Lipschitz condition (with respect to $y$) on $D$, with Lipschitz constant $L\ge0$, if

$$|f(x,y_1)-f(x,y_2)|\le L\,|y_1-y_2|\qquad\text{whenever }(x,y_1),(x,y_2)\in D.$$

Only the second variable is involved: the two points have the same $x$. We say $f$ is locally Lipschitz in $y$ on an open set $D$ if every point of $D$ has a neighbourhood (a rectangle) on which $f$ is Lipschitz in $y$, with a constant that may depend on the neighbourhood.

📐 Theorem 1.2 — A sufficient condition via $\partial f/\partial y$

Let $D$ be convex in $y$ (whenever $(x,y_1),(x,y_2)\in D$, the vertical segment joining them lies in $D$ — e.g. a rectangle or a vertical strip).

  1. If $\partial f/\partial y$ exists on $D$ and $\left|\partial f/\partial y\right|\le L$ there, then $f$ is Lipschitz in $y$ on $D$ with constant $L$.
  2. Conversely, if $f$ is Lipschitz in $y$ with constant $L$ and $\partial f/\partial y$ exists at an interior point, then $|\partial f/\partial y|\le L$ there.
  3. If $D$ is open and $\partial f/\partial y$ is continuous on $D$, then $f$ is locally Lipschitz in $y$ on $D$.

Proof. (1) Fix $x$ and $y_1\lt y_2$ with $(x,y_1),(x,y_2)\in D$. The function $g(s)=f(x,s)$ is differentiable on $[y_1,y_2]$ (the segment lies in $D$), so by the mean value theorem $f(x,y_2)-f(x,y_1)=\tfrac{\partial f}{\partial y}(x,\eta)(y_2-y_1)$ for some $\eta$ between $y_1$ and $y_2$; taking absolute values gives the bound with constant $L$. (2) $\left|\tfrac{f(x,y+k)-f(x,y)}{k}\right|\le L$ for all small $k\ne0$; let $k\to0$. (3) Around each point of $D$ choose a closed rectangle $R\subset D$. The continuous function $|\partial f/\partial y|$ attains a maximum $L_R$ on the compact set $R$, and $R$ is convex, so (1) applies on $R$. ∎

The theorem is only a sufficient test. The function $f(x,y)=|y|$ is Lipschitz with $L=1$ on all of $\mathbb{R}^2$ (by the reverse triangle inequality $\big||y_1|-|y_2|\big|\le|y_1-y_2|$) although $\partial f/\partial y$ does not exist on the line $y=0$. And a Lipschitz constant must work for all pairs of points: $f=y^2$ is Lipschitz on every bounded rectangle but not on a strip $[0,1]\times\mathbb{R}$, since $|y_1^2-y_2^2|/|y_1-y_2|=|y_1+y_2|$ is unbounded there.

✏️ Example 2 — Lipschitz or not?
Decide whether $f$ is Lipschitz in $y$ on the given set, and if so find the best constant: (a) $f=y^2$ on $R=[-1,1]\times[-2,2]$ and on $[0,1]\times\mathbb{R}$; (b) $f=y^{1/3}$ (real cube root) on $[-1,1]\times[-1,1]$; (c) $f=x|y|$ on $[-3,3]\times\mathbb{R}$; (d) $f=\dfrac{x^2}{1+y^2}$ on $[-2,2]\times\mathbb{R}$.
(a) On $R$: $|\partial f/\partial y|=2|y|\le4$, so $L=4$, and $4$ is best since $|y_1+y_2|$ approaches $4$ for $y_1,y_2$ near $2$. On the strip $[0,1]\times\mathbb{R}$: no Lipschitz constant, because $|f(x,y_1)-f(x,y_2)|/|y_1-y_2|=|y_1+y_2|$ is unbounded.
(b) Not Lipschitz Take $y_1=k\gt0$, $y_2=0$: $\dfrac{|k^{1/3}-0|}{|k-0|}=k^{-2/3}\to\infty$ as $k\to0^+$. No constant $L$ can work on any rectangle that contains a point of the $x$-axis in its interior or on its edge, although $f$ is continuous everywhere.
(c) $L=3$ $\big|x|y_1|-x|y_2|\big|=|x|\,\big||y_1|-|y_2|\big|\le3|y_1-y_2|$, with equality at $x=3$, $y_1,y_2\gt0$. Lipschitz although $\partial f/\partial y$ fails to exist on $y=0$ (for $x\ne0$).
(d) $L=\tfrac{3\sqrt3}{2}$ $\dfrac{\partial f}{\partial y}=-\dfrac{2x^2y}{(1+y^2)^2}$. With $g(y)=\dfrac{y}{(1+y^2)^2}$, $g'(y)=\dfrac{1-3y^2}{(1+y^2)^3}$, so $|g|$ is largest at $y=\pm\tfrac1{\sqrt3}$, where $|g|=\tfrac{1/\sqrt3}{16/9}=\tfrac{9}{16\sqrt3}$. Hence $\left|\tfrac{\partial f}{\partial y}\right|\le2\cdot4\cdot\tfrac{9}{16\sqrt3}=\tfrac{3\sqrt3}{2}\approx2.598$ on the whole strip, which is convex in $y$. By Theorem 1.6 below, every IVP for $y'=x^2/(1+y^2)$ with $x_0\in[-2,2]$ has a unique solution on all of $[-2,2]$.

Picard Iterates and the Picard–Lindelöf Theorem

Theorem 1.1 turns the IVP into a fixed-point problem: $\varphi=T\varphi$ for the operator $(T\varphi)(x)=y_0+\int_{x_0}^x f(t,\varphi(t))\,dt$. The most natural way to look for a fixed point is to iterate.

📘 Definition 1.3 — Picard iterates (successive approximations)
The Picard iterates of the IVP $y'=f(x,y)$, $y(x_0)=y_0$ are the functions $$\varphi_0(x)=y_0,\qquad \varphi_{n+1}(x)=y_0+\int_{x_0}^{x}f\big(t,\varphi_n(t)\big)\,dt\quad(n\ge0),$$ defined on any interval on which the graphs of $\varphi_0,\dots,\varphi_n$ stay in the domain of $f$.
📐 Theorem 1.3 — Picard–Lindelöf (Cauchy–Lipschitz) existence and uniqueness theorem

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

  1. continuous on $R$, with $M=\max_R|f|$, and
  2. Lipschitz in $y$ on $R$ with constant $L$.

Put $h=\min\left(a,\dfrac bM\right)$ (and $h=a$ if $M=0$), $I=[x_0-h,x_0+h]$. Then the IVP $y'=f(x,y)$, $y(x_0)=y_0$ has exactly one solution $\varphi$ on $I$. The Picard iterates are defined on $I$, have graphs in $R$, and converge to $\varphi$ uniformly on $I$, with

$$|\varphi(x)-\varphi_n(x)|\le\frac{M}{L}\,\frac{(Lh)^{n+1}}{(n+1)!}\,e^{Lh}\qquad(x\in I,\ L\gt0).$$

Moreover, any solution of the IVP on any interval $J\subseteq I$ containing $x_0$ coincides with $\varphi$ on $J$.

Proof. Step 1: the iterates stay in $R$. We show by induction that each $\varphi_n$ is continuous on $I$ with $|\varphi_n(x)-y_0|\le b$. This is true for $\varphi_0$. If it holds for $\varphi_n$, then $t\mapsto f(t,\varphi_n(t))$ is continuous on $I$ and bounded by $M$, so $\varphi_{n+1}$ is defined, continuous (indeed $C^1$), and $|\varphi_{n+1}(x)-y_0|\le M|x-x_0|\le Mh\le b$. This is exactly where $h\le b/M$ is needed: the graph of $\varphi_{n+1}$ cannot escape through the top or bottom of $R$ before $|x-x_0|$ reaches $h$.

Step 2: the successive differences shrink factorially. We claim $$|\varphi_{k+1}(x)-\varphi_k(x)|\le\frac{ML^k|x-x_0|^{k+1}}{(k+1)!}\qquad(x\in I,\ k\ge0).$$ For $k=0$ this is $|\varphi_1-y_0|\le M|x-x_0|$ from Step 1. If it holds for $k-1$, then for $x\ge x_0$ (the other side is symmetric), by the Lipschitz condition, $$|\varphi_{k+1}(x)-\varphi_k(x)|\le\int_{x_0}^x\big|f(t,\varphi_k)-f(t,\varphi_{k-1})\big|dt\le L\int_{x_0}^x\frac{ML^{k-1}(t-x_0)^k}{k!}dt=\frac{ML^k(x-x_0)^{k+1}}{(k+1)!}.$$

Step 3: uniform convergence. On $I$ the $k$-th difference is at most $\dfrac{M}{L}\dfrac{(Lh)^{k+1}}{(k+1)!}$, and $\sum_k\tfrac{(Lh)^{k+1}}{(k+1)!}=e^{Lh}-1\lt\infty$. By the Weierstrass M-test, $\varphi_n=\varphi_0+\sum_{k=0}^{n-1}(\varphi_{k+1}-\varphi_k)$ converges uniformly on $I$ to a continuous function $\varphi$, and $|\varphi(x)-y_0|\le b$ because each $\varphi_n$ satisfies this. The tail estimate is $$|\varphi-\varphi_n|\le\frac ML\sum_{j\ge n+1}\frac{(Lh)^j}{j!}\le\frac ML\frac{(Lh)^{n+1}}{(n+1)!}\sum_{i\ge0}\frac{(Lh)^i}{i!},$$ using $(n+1+i)!\ge(n+1)!\,i!$; this is the stated error bound.

Step 4: the limit is a solution. $|f(t,\varphi_n(t))-f(t,\varphi(t))|$ is at most $L|\varphi_n(t)-\varphi(t)|\le L\sup_I|\varphi_n-\varphi|$, which tends to $0$, uniformly in $t$, so $\int_{x_0}^xf(t,\varphi_n)\,dt\to\int_{x_0}^xf(t,\varphi)\,dt$. Letting $n\to\infty$ in the definition of $\varphi_{n+1}$ gives $\varphi=y_0+\int_{x_0}^xf(t,\varphi)\,dt$, and Theorem 1.1 shows that $\varphi$ solves the IVP.

Step 5: uniqueness. Let $\psi$ be a solution on $J\subseteq I$. First, the graph of $\psi$ stays in $R$. If not, by continuity there is a first point $x_1\gt x_0$ (say) of $J$ at which $|\psi(x_1)-y_0|=b$, followed by points where the graph is outside $R$; on $[x_0,x_1]$ the graph is in $R$, so $|\psi'|\le M$ there and $b=|\psi(x_1)-y_0|\le M(x_1-x_0)$. Hence $x_1-x_0\ge b/M\ge h$, so $x_1=x_0+h$ is the right end of $I$ and there are no later points of $J$ — a contradiction. Now $u=|\varphi-\psi|$ is continuous, bounded by $2b$ on $J$, and $u(x)\le L\int_{x_0}^xu(t)\,dt$. Inserting the bound $u\le2b$ into the right-hand side repeatedly gives $u(x)\le2b\,\dfrac{L^n|x-x_0|^n}{n!}$ for every $n$, and the right-hand side tends to $0$. So $\psi=\varphi$ on $J$. ∎

Three features of the proof are worth remembering. The Lipschitz condition is used twice — for convergence (Step 2) and for uniqueness (Step 5). The bound $M$ only controls where the iterates can go, so it limits the interval through $h=\min(a,b/M)$, not the convergence. And the iterates converge factorially fast, like the partial sums of $e^{Lh}$: Picard iteration is a contraction-mapping argument in disguise (with the norm $\sup_I e^{-2L|x-x_0|}|\varphi(x)|$ the operator $T$ is a contraction on the whole of $I$; the Banach fixed-point theorem is proved in the CSIR NET Unit 1 course, in its chapter on metric spaces).

✏️ Example 3 — Picard iterates that we can sum
Compute the Picard iterates $\varphi_1,\varphi_2,\varphi_3$ for $y'=2x(1+y)$, $y(0)=0$, guess $\varphi_n$, and identify the limit.
Iterate $\varphi_0=0$. $\varphi_1=\int_0^x2t\,dt=x^2$. $\varphi_2=\int_0^x2t(1+t^2)\,dt=x^2+\tfrac{x^4}{2}$. $\varphi_3=\int_0^x2t\left(1+t^2+\tfrac{t^4}{2}\right)dt=x^2+\tfrac{x^4}{2}+\tfrac{x^6}{6}$.
Pattern By induction, $\varphi_n=\sum_{k=1}^n\dfrac{x^{2k}}{k!}$: if this holds for $n$, then $\varphi_{n+1}=\int_0^x\left(2t+\sum_{k=1}^n\tfrac{2t^{2k+1}}{k!}\right)dt=x^2+\sum_{k=1}^n\tfrac{x^{2k+2}}{(k+1)!}$.
Limit $\varphi_n(x)\to\sum_{k\ge1}\tfrac{(x^2)^k}{k!}=e^{x^2}-1$, uniformly on bounded intervals. Check: $y=e^{x^2}-1$ gives $y'=2xe^{x^2}=2x(1+y)$, $y(0)=0$.
Remark Here $f=2x(1+y)$ is Lipschitz on every strip $[-T,T]\times\mathbb{R}$ (with $L=2T$), so the iterates converge on every $[-T,T]$ — not just on the small interval $h$ of Theorem 1.3. Theorem 1.6 explains why.
Interactive 1 — Picard Iterates Closing In on a Solution

The blue curve is the exact solution, the red curve the current Picard iterate $\varphi_n$, and grey curves the earlier iterates. The green band is the interval $|x|\le h$ guaranteed by Theorem 1.3 for the rectangle named in the readout. The iterates are computed numerically (trapezoid rule, step $0.005$). Press Play or Next; in the last case watch what happens beyond $x=1$.

iterate n—
rectangle → h—
max |φₙ − y| on |x| ≤ h—
max |φₙ − φₙ₋₁| on the window—
✏️ Example 4 — The guaranteed interval versus the true one
For $y'=1+y^2$, $y(0)=0$, use Theorem 1.3 on rectangles $|x|\le a$, $|y|\le b$ to find the largest interval of existence the theorem can guarantee, and compare it with the actual maximal interval.
Constants On the rectangle, $M=\max(1+y^2)=1+b^2$ and $L=\max|2y|=2b$. So $h=\min\left(a,\dfrac{b}{1+b^2}\right)$.
Optimise $\dfrac{d}{db}\dfrac{b}{1+b^2}=\dfrac{1-b^2}{(1+b^2)^2}$ vanishes at $b=1$, where $\tfrac{b}{1+b^2}=\tfrac12$. Taking any $a\ge\tfrac12$, the best guaranteed interval is $|x|\le\tfrac12$.
Truth Separating variables, $\arctan y=x$, so $y=\tan x$, which exists exactly on $\left(-\tfrac\pi2,\tfrac\pi2\right)$ and blows up at the ends.
Lesson Theorem 1.3 is a local theorem with a conservative interval: $\tfrac12$ versus $\tfrac\pi2\approx1.571$. Restarting it at the end point $(\tfrac12,\tan\tfrac12)$ extends the solution further; the extension theorem (Theorem 1.5) says that this process continues until the solution leaves every compact subset of the domain.
📐 Theorem 1.4 — Peano's existence theorem
If $f$ is merely continuous on $R=\{|x-x_0|\le a,\ |y-y_0|\le b\}$ with $M=\max_R|f|\gt0$, then the IVP $y'=f(x,y)$, $y(x_0)=y_0$ has at least one solution on $[x_0-h,x_0+h]$, $h=\min(a,b/M)$. Uniqueness may fail.

Proof (sketch). (1) Euler polygons. For each $n$ divide $[x_0,x_0+h]$ into $n$ equal steps and join the points $(x_k,y_k)$ with $y_{k+1}=y_k+\tfrac hn f(x_k,y_k)$ by segments. Each polygon $\psi_n$ has slopes bounded by $M$, so it stays in $R$ (as in Step 1 above) and $|\psi_n(x)-\psi_n(x')|\le M|x-x'|$. (2) Compactness. The family $\{\psi_n\}$ is uniformly bounded and equicontinuous, so by the Arzelà–Ascoli theorem a subsequence converges uniformly to a continuous $\varphi$. (3) The limit solves the integral equation. Each polygon satisfies $\psi_n(x)=y_0+\int_{x_0}^xf(t,\psi_n(t))\,dt+\varepsilon_n(x)$ with $\sup|\varepsilon_n|\to0$, by uniform continuity of $f$ on $R$; passing to the limit along the subsequence gives the integral equation for $\varphi$, and Theorem 1.1 finishes. Without the Lipschitz condition different subsequences can converge to different solutions. ∎

Non-Uniqueness, Maximal Solutions and Blow-Up

✏️ Example 5 — Infinitely many solutions, and uniqueness that is only local
(a) Find all solutions of $y'=3y^{2/3}$, $y(0)=0$ on $\mathbb{R}$ (here $y^{2/3}=(y^{1/3})^2\ge0$). (b) Show that the IVP $y'=3y^{2/3}$, $y(0)=1$ has a unique solution near $x=0$, but that its continuation to $\mathbb{R}$ is not unique.
Building blocks $y=(x-c)^3$ is a solution on $\mathbb{R}$: $y'=3(x-c)^2=3\left((x-c)^3\right)^{2/3}$. So is $y\equiv0$. Since $f=3y^{2/3}\ge0$, every solution is non-decreasing.
(a) Glue For any $d\le0\le c$ put $$y=\begin{cases}(x-d)^3,&x\lt d,\\0,&d\le x\le c,\\(x-c)^3,&x\gt c.\end{cases}$$ At $x=c$ both one-sided derivatives are $0=3\cdot0^{2/3}$ (and the same at $d$), so $y$ is differentiable and solves the IVP; $d=-\infty$ or $c=+\infty$ (no cubic piece on that side) is allowed. Conversely, a solution with $y(0)=0$ vanishes on a closed interval $[d,c]\ni0$ (the zero set of a non-decreasing function is an interval); on $(c,\infty)$ it is positive, and there $y^{-2/3}y'=3$ integrates to $y^{1/3}=x-c$. The negative side is the same. So these are all the solutions — uncountably many.
(b) Local uniqueness On the rectangle $|x|\le\tfrac14$, $|y-1|\le\tfrac12$ we have $\partial f/\partial y=2y^{-1/3}$, bounded by $2\cdot(1/2)^{-1/3}\lt2.6$; Theorem 1.2 and Theorem 1.3 give a unique solution near $0$, namely $y=(x+1)^3$.
(b) Global failure Followed backwards, $y=(x+1)^3$ reaches $0$ at $x=-1$, where $f$ is not Lipschitz. For $x\lt-1$ the solution may stay at $0$ until any $d\le-1$ and then continue as $(x-d)^3$. So the IVP has infinitely many solutions on $\mathbb{R}$, all equal to $(x+1)^3$ on $[-1,\infty)$. Uniqueness is a local property, lost wherever the graph meets the set where $f$ fails to be Lipschitz.

The same gluing works for $y'=y^{1/3}$ (blocks $y=\pm\left(\tfrac{2(x-c)}{3}\right)^{3/2}$ for $x\ge c$) and for $y'=\sqrt{|y|}$ (blocks $y=\tfrac{(x-c)^2}4$ for $x\ge c$). More generally $y'=|y|^\alpha$, $y(0)=0$ has infinitely many solutions for $0\lt\alpha\lt1$ and only $y\equiv0$ for $\alpha\ge1$.

📘 Definition 1.4 — Extension, maximal solution, maximal interval
A solution $\psi$ on $J$ extends a solution $\varphi$ on $I$ if $I\subseteq J$ and $\psi=\varphi$ on $I$. A solution is maximal (non-extendable) if it has no extension to a strictly larger interval; its interval of definition, written $(\omega_-,\omega_+)$, is the maximal interval of existence. A solution defined on all of $\mathbb{R}$ (or on the whole $x$-range of $D$) is called global.
📐 Theorem 1.5 — Maximal solutions and the escape (blow-up) alternative

Let $D\subseteq\mathbb{R}^2$ be open and $f:D\to\mathbb{R}$ continuous and locally Lipschitz in $y$. Then for every $(x_0,y_0)\in D$:

  1. the IVP has a unique maximal solution $\varphi$, and its maximal interval $(\omega_-,\omega_+)$ is open; every other solution is a restriction of $\varphi$;
  2. as $x\to\omega_+$ the point $(x,\varphi(x))$ leaves every compact subset $K$ of $D$: there is $x_K\lt\omega_+$ with $(x,\varphi(x))\notin K$ for $x_K\lt x\lt\omega_+$ (and similarly at $\omega_-$).

In particular, if $D=(\alpha,\beta)\times\mathbb{R}$, then either $\omega_+=\beta$ or $|\varphi(x)|\to\infty$ as $x\to\omega_+^-$: a solution on a strip can only end by blowing up.

Proof (sketch). (1) If two solutions $\varphi_1$ on $I_1$ and $\varphi_2$ on $I_2$ agree at one point $x_1\in I_1\cap I_2$, the set $\{x\in I_1\cap I_2:\varphi_1(x)=\varphi_2(x)\}$ is closed (continuity) and open (apply Theorem 1.3 on a small rectangle around each point where they agree), hence it is the whole interval $I_1\cap I_2$. So all solutions of the IVP agree where they overlap, and their union defines one solution on the union of their intervals — the maximal solution. The interval is open, because a solution on $[x_0,\omega_+]$ could be extended past $\omega_+$ by Theorem 1.3 at $(\omega_+,\varphi(\omega_+))\in D$.

(2) Let $K\subset D$ be compact. Choose $\delta\gt0$ so that the compact set $K_\delta$ of points within distance $\delta$ (in the max-norm) of $K$ lies in $D$, and let $M$ bound $|f|$ on $K_\delta$; $f$ is Lipschitz in $y$ on $K_\delta$ (cover it by finitely many rectangles). For every $(\xi,\eta)\in K$, Theorem 1.3 on the square of half-side $\delta$ centred there gives a solution through $(\xi,\eta)$ on $[\xi,\xi+h]$ with the same $h=\min(\delta,\delta/M)$. If there were points $x_k\to\omega_+$ with $(x_k,\varphi(x_k))\in K$, pick one with $\omega_+-x_k\lt h$: the solution through $(x_k,\varphi(x_k))$ would extend $\varphi$ beyond $\omega_+$. Contradiction. For the strip, if $\omega_+\lt\beta$ apply this with $K=[x_0,\omega_+]\times[-N,N]$ for every $N$. ∎

✏️ Example 6 — Blow-up for $y'=y^2$
Find the maximal solution of $y'=y^2$, $y(0)=y_0$ for every $y_0\in\mathbb{R}$, and compare its life span with the interval guaranteed by Theorem 1.3.
$y_0=0$ $y\equiv0$ is a solution on $\mathbb{R}$; by uniqueness ($f=y^2$ is $C^1$, hence locally Lipschitz) it is the maximal one, and no other solution ever touches the axis.
$y_0\ne0$ Then $y$ never vanishes, and $-\left(\tfrac1y\right)'=\tfrac{y'}{y^2}=1$, so $\tfrac1y=\tfrac1{y_0}-x$: $$y=\frac{y_0}{1-y_0x}.$$ For $y_0\gt0$ the maximal interval is $\left(-\infty,\tfrac1{y_0}\right)$; for $y_0\lt0$ it is $\left(\tfrac1{y_0},\infty\right)$. In both cases $|y|\to\infty$ at the finite end, exactly as Theorem 1.5 predicts for the strip $\mathbb{R}\times\mathbb{R}$.
Theorem 1.3 Take $y_0\gt0$, $|y-y_0|\le b$, any $a$: $M=(y_0+b)^2$, $h=\min\left(a,\tfrac{b}{(y_0+b)^2}\right)$. The function $b\mapsto\tfrac{b}{(y_0+b)^2}$ is maximal at $b=y_0$, giving $h=\tfrac1{4y_0}$ — a quarter of the true forward life span $\tfrac1{y_0}$.
Moral Smoothness of $f$ does not prevent blow-up: superlinear growth in $y$ does it. Linear growth does not (Theorem 1.6).
Interactive 2 — Uniqueness Lost, Solutions That Die: Slope Fields and Maximal Intervals

Short segments show the slope field $f(x,y)$. In the first mode every red curve leaves the axis at a different point $c$ and all of them solve the same IVP $y(0)=0$ — the Lipschitz condition fails on $y=0$. In the other modes the blue solution through $(0,y_0)$ is traced until it blows up at the dashed asymptote $x=\omega_+$; the green band is the best interval Theorem 1.3 can guarantee. Move the slider.

—
parameter—
maximal interval—
Picard guarantee—
Lipschitz near the curve?—
📐 Theorem 1.6 — Global existence under a global Lipschitz condition
Let $f$ be continuous on the strip $S=[\alpha,\beta]\times\mathbb{R}$ and Lipschitz in $y$ on $S$ with constant $L$. Then for every $x_0\in[\alpha,\beta]$ and $y_0\in\mathbb{R}$ the IVP $y'=f(x,y)$, $y(x_0)=y_0$ has a unique solution on the whole interval $[\alpha,\beta]$, and the Picard iterates converge to it uniformly there. In particular, the linear equation $y'=p(x)y+q(x)$ with $p,q$ continuous on an interval $J$ has, for every $x_0\in J$, a unique solution defined on all of $J$.

Proof. Now the iterates need not stay in any rectangle — $f$ is defined for every $y$ — so they are defined and continuous on all of $[\alpha,\beta]$. Let $M_0=\max_{[\alpha,\beta]}|f(t,y_0)|$. Then $|\varphi_1(x)-\varphi_0(x)|\le M_0|x-x_0|$, and the induction of Step 2 of Theorem 1.3 (which used only the Lipschitz condition) gives $|\varphi_{k+1}-\varphi_k|\le$ $M_0L^k|x-x_0|^{k+1}/(k+1)!$ on $[\alpha,\beta]$. Steps 3–5 go through verbatim with $h$ replaced by $\beta-\alpha$ (uniqueness now needs no argument about staying in $R$). For the linear equation, the difference of the right-hand sides at $y_1$ and at $y_2$ is $p(x)(y_1-y_2)$, of absolute value at most $\left(\max_{[\alpha,\beta]}|p|\right)|y_1-y_2|$ on every compact $[\alpha,\beta]\subseteq J$; the solutions on an increasing sequence of compact intervals exhausting $J$ agree on overlaps by uniqueness. ∎

This is the reason linear equations never blow up: a linear equation can be singular only where its coefficients are, and the solution lives on the whole interval on which $p$ and $q$ are continuous. Chapter 2 rests on this fact.

Gronwall's Inequality and Continuous Dependence

A physical law is useful only if small errors in the data produce small errors in the prediction. The tool for proving this is an inequality that turns an integral inequality into an explicit exponential bound.

📐 Theorem 1.7 — Gronwall's inequality
Let $u$ and $k$ be continuous on $[x_0,x_1]$ with $k\ge0$, and let $C\in\mathbb{R}$. If $$u(x)\le C+\int_{x_0}^{x}k(t)\,u(t)\,dt\qquad(x_0\le x\le x_1),$$ then $$u(x)\le C\exp\!\left(\int_{x_0}^{x}k(t)\,dt\right)\qquad(x_0\le x\le x_1).$$ In particular, if $u\ge0$ and $u(x)\le L\int_{x_0}^xu$, then $u\equiv0$.

Proof. Put $U(x)=C+\int_{x_0}^xk(t)u(t)\,dt$, so $u\le U$ and $U$ is $C^1$ with $U'=ku\le kU$ (here $k\ge0$ is essential). With $K(x)=\int_{x_0}^xk$, $$\left(U e^{-K}\right)'=(U'-kU)e^{-K}\le0,$$ so $U(x)e^{-K(x)}\le U(x_0)=C$. Hence $u(x)\le U(x)\le Ce^{K(x)}$. The last statement is the case $C=0$, $k=L$. ∎

📐 Theorem 1.8 — Continuous dependence on initial data and on the equation
Let $f,g$ be continuous on a region $D$, with $f$ Lipschitz in $y$ (constant $L\gt0$) and $|f-g|\le\varepsilon$ on $D$. Let $y$ solve $y'=f(x,y)$, $y(x_0)=y_0$, and $z$ solve $z'=g(x,z)$, $z(x_0)=z_0$, both on $[x_0,x_1]$ with graphs in $D$. Then for $x_0\le x\le x_1$ $$|y(x)-z(x)|\le|y_0-z_0|\,e^{L(x-x_0)}+\frac{\varepsilon}{L}\left(e^{L(x-x_0)}-1\right).$$ The same holds for $x\le x_0$ with $|x-x_0|$ in place of $x-x_0$.

Proof. Subtract the integral equations: $$y(x)-z(x)=y_0-z_0+\int_{x_0}^x\Big(\big[f(t,y)-f(t,z)\big]+\big[f(t,z)-g(t,z)\big]\Big)dt.$$ Hence $w=|y-z|$ satisfies $w(x)\le\delta+\int_{x_0}^x(Lw+\varepsilon)\,dt$, with $\delta=|y_0-z_0|$. Put $u=w+\varepsilon/L$: then $u(x)\le\left(\delta+\tfrac\varepsilon L\right)+\int_{x_0}^xLu\,dt$, and Gronwall gives $u\le\left(\delta+\tfrac\varepsilon L\right)e^{L(x-x_0)}$, which is the claim. ∎

With $\varepsilon=0$ and $\delta=0$ this is once more the uniqueness theorem. The factor $e^{L(x-x_0)}$ is honest: for $y'=y$ two solutions starting $\delta$ apart are $\delta e^{x}$ apart at time $x$. Continuous dependence on a finite interval is not the same as stability for all time — that question returns in §1.3.

Tools for Solving: Exact Equations, Integrating Factors, Linear and Bernoulli Equations

The existence theory tells us a solution exists; for many equations met in practice the classical methods find it. We collect them briefly — CSIR questions often combine a solution method with an existence question (find the solution, then its maximal interval).

📘 Definition 1.5 — Exact equation and integrating factor
The equation $M(x,y)\,dx+N(x,y)\,dy=0$ (meaning $M+Ny'=0$) is exact on a region $D$ if there is a $C^1$ function $\Phi$ on $D$ with $\Phi_x=M$ and $\Phi_y=N$; then its solutions satisfy $\Phi(x,y)=c$. A nonzero function $\mu(x,y)$ is an integrating factor if $\mu M\,dx+\mu N\,dy=0$ is exact.
📐 Theorem 1.9 — Test for exactness
Let $M,N$ be $C^1$ on an open rectangle $D$ (more generally on a simply connected region). Then $M\,dx+N\,dy=0$ is exact on $D$ if and only if $\dfrac{\partial M}{\partial y}=\dfrac{\partial N}{\partial x}$ on $D$. In that case, for any $(x_0,y_0)\in D$, $$\Phi(x,y)=\int_{x_0}^{x}M(t,y_0)\,dt+\int_{y_0}^{y}N(x,s)\,ds.$$

Proof. If $\Phi_x=M$, $\Phi_y=N$ with $M,N\in C^1$, then $\Phi$ is $C^2$ and by Schwarz's theorem $M_y=\Phi_{xy}=\Phi_{yx}=N_x$. Conversely, with $\Phi$ as displayed (the segments used stay in the rectangle), $\Phi_y=N(x,y)$ by the fundamental theorem of calculus, and differentiating under the integral sign, $$\Phi_x=M(x,y_0)+\int_{y_0}^yN_x(x,s)\,ds=M(x,y_0)+\int_{y_0}^yM_y(x,s)\,ds=M(x,y).$$ If $y(x)$ solves $M+Ny'=0$ then $\tfrac{d}{dx}\Phi(x,y(x))=\Phi_x+\Phi_yy'=0$, so $\Phi$ is constant along solutions. ∎

📐 Theorem 1.10 — Integrating factors, linear and Bernoulli equations
  1. If $\dfrac{M_y-N_x}{N}=g(x)$ depends on $x$ alone, then $\mu=e^{\int g\,dx}$ is an integrating factor; if $\dfrac{N_x-M_y}{M}=h(y)$ depends on $y$ alone, then $\mu=e^{\int h\,dy}$ is one.
  2. The linear equation $y'+P(x)y=Q(x)$ ($P,Q$ continuous on $J$) has the general solution $$y=e^{-\int P}\left(\int Qe^{\int P}dx+C\right)\quad\text{on }J.$$
  3. The Bernoulli equation $y'+P(x)y=Q(x)y^n$ ($n\ne0,1$) becomes linear for $v=y^{1-n}$: $\ v'+(1-n)Pv=(1-n)Q$. For $n\gt0$, $y\equiv0$ is an extra solution, and for $0\lt n\lt1$ the IVP with $y(x_0)=0$ is in general not unique.

Proof. (1) For $\mu=\mu(x)$, exactness of $\mu M\,dx+\mu N\,dy$ means $\mu M_y=\mu'N+\mu N_x$, i.e. $\mu'/\mu=(M_y-N_x)/N$, which is solvable by $\mu=e^{\int g}$ exactly when the right side depends only on $x$; the $y$-case is symmetric. (2) Multiply by $\mu=e^{\int P}$: $(\mu y)'=\mu y'+\mu Py=\mu Q$; integrate. (3) For $y\ne0$, $v'=(1-n)y^{-n}y'$, and $y^{-n}y'=y^{-n}(Qy^n-Py)=Q-Pv$, so $v'=(1-n)(Q-Pv)$. For $0\lt n\lt1$ the right side $Qy^n-Py$ is not Lipschitz at $y=0$, which is how the gluing of Example 5 becomes possible. ∎

✏️ Example 7 — A Bernoulli IVP and its maximal interval
Solve $y'+\dfrac{y}{x}=xy^2$, $y(1)=1$, and find the maximal interval of existence.
Domain $f(x,y)=xy^2-\tfrac yx$ is $C^1$ on the open half-planes $x\gt0$ and $x\lt0$; since $x_0=1$, we work in $D=(0,\infty)\times\mathbb{R}$, where Theorem 1.5 applies.
Substitute Bernoulli with $n=2$: $v=y^{-1}$ satisfies $v'-\tfrac vx=-x$. Integrating factor $\tfrac1x$: $\left(\tfrac vx\right)'=-1$, so $\tfrac vx=-x+C$, and $v(1)=1$ gives $C=2$: $v=2x-x^2$, $$y=\frac{1}{x(2-x)}.$$
Maximal interval The formula is finite on $(0,2)$, and $y\to+\infty$ as $x\to0^+$ and as $x\to2^-$. By Theorem 1.5 the maximal interval is $(0,2)$: at $x=2$ the solution blows up inside $D$; at $x=0$ it reaches the boundary of $D$ (and blows up as well). The formula also makes sense for $x\gt2$, but that branch is a different solution, not an extension.
Check $y=(2x-x^2)^{-1}$ gives $y'=-\dfrac{2-2x}{(2x-x^2)^2}=\dfrac{2x-2}{x^2(2-x)^2}$. On the other side, $xy^2=\dfrac{x}{x^2(2-x)^2}$ and $\dfrac yx=\dfrac{1}{x^2(2-x)}=\dfrac{2-x}{x^2(2-x)^2}$, so $xy^2-\dfrac yx=\dfrac{x-(2-x)}{x^2(2-x)^2}=\dfrac{2x-2}{x^2(2-x)^2}$. Equal, and $y(1)=1$.
🧭 Problem-solving strategy — existence and uniqueness questions
  1. Locate the bad set. Find where $f$ is discontinuous or where $\partial f/\partial y$ fails to exist or is unbounded (typically $y^\alpha$ with $0\lt\alpha\lt1$ at $y=0$, or a denominator vanishing).
  2. Away from it, $f$ is locally Lipschitz: through every point there is exactly one solution (Theorems 1.2, 1.3, 1.5).
  3. On it, test uniqueness directly: try to glue a constant solution to a nonconstant one (Example 5). If $f(y_0)\ne0$ for an autonomous equation, separation of variables usually restores uniqueness even without a Lipschitz condition.
  4. Interval questions. For "guaranteed by Picard" compute $M$ on the given rectangle and $h=\min(a,b/M)$. For "maximal interval" solve explicitly and use Theorem 1.5 (blow-up or reaching $\partial D$); for linear or globally Lipschitz $f$ use Theorem 1.6.
  5. Dependence questions. Use Theorem 1.8 with the Lipschitz constant on the relevant strip.
⚠️ Counterexamples and pitfalls
  1. Continuity does not give uniqueness. $y'=3y^{2/3}$, $y(0)=0$: $y\equiv0$, $y=x^3$ and uncountably many glued solutions (Example 5). Likewise $y'=\sqrt{|y|}$ and $y'=y^{1/3}$.
  2. Without continuity there may be no solution. $f(x,y)=-1$ for $y\ge0$, $f=1$ for $y\lt0$, $y(0)=0$: a solution would have $y'(0)=-1$, so $y\lt0$ just after $0$, where $y'=1$ forces $y$ to increase from $0$ — a contradiction. No solution on any $[0,\delta)$.
  3. Lipschitz is sufficient, not necessary, for uniqueness. $y'=1+y^{2/3}$, $y(0)=0$ has a unique solution (separate: $\int_0^y\frac{ds}{1+s^{2/3}}=x$ with a strictly increasing left side), although $f$ is not Lipschitz at $y=0$.
  4. Lipschitz does not need differentiability. $|y|$, $x|y|$, $|\sin y|$ are Lipschitz but not differentiable in $y$ everywhere.
  5. A bounded $\partial f/\partial y$ needs a domain convex in $y$. On a set made of two separated horizontal bands, the mean value theorem cannot be applied across the gap.
  6. Smooth $f$ can blow up. $y'=y^2$ and $y'=1+y^2$: $f$ is a polynomial, yet the solutions live on finite intervals. "Locally Lipschitz" gives local existence only.
  7. $h=\min(a,b/M)$ is a guarantee, not the answer. For $y'=1+y^2$, $y(0)=0$ the best guarantee is $\tfrac12$, the truth $\tfrac\pi2$; for $y'=y$ every rectangle gives $h\le1$, the truth is $\mathbb{R}$.
  8. Picard iterates may converge to one of several solutions. For $y'=3y^{2/3}$, $y(0)=0$ all iterates are $0$: they converge — to just one of infinitely many solutions. Convergence of iterates does not prove uniqueness.
  9. Uniqueness is local. $y'=3y^{2/3}$, $y(0)=1$ is unique near $0$ but has infinitely many global solutions (Example 5(b)).
  10. Gronwall needs $k\ge0$. With $C=1$, $k=-1$ on $[0,1]$: $u\equiv\tfrac12$ satisfies $u\le1-\int_0^xu\,dt=1-\tfrac x2$, yet the "conclusion" $u\le e^{-x}$ fails at $x=1$ ($e^{-1}\approx0.37\lt\tfrac12$). The step $ku\le kU$ of the proof reverses when $k\lt0$.

Exercises

I
Integral equations and Lipschitz conditions.
  • 1 Write $y'=x^2+y^2$, $y(0)=1$ as an integral equation and compute the Picard iterates $\varphi_1$ and $\varphi_2$.
  • 2 Find all continuous $y$ on $(-\infty,1)$ with $y(x)=1+\int_0^xy(t)^2\,dt$.
  • 3 Is $f$ Lipschitz in $y$ on the given set? Give the best constant if so. (a) $xy^2$ on $[0,1]\times[-3,3]$; (b) $e^{-x}\sin y$ on $[0,\infty)\times\mathbb{R}$; (c) $y^{2/3}$ on $[0,1]\times[-1,1]$; (d) $x|y|$ on $[-2,2]\times\mathbb{R}$.
  • 4 Show that $f(x,y)=\sqrt{|y|}$ is Lipschitz in $y$ on $\{(x,y):|y|\ge\delta\}$ for every $\delta\gt0$, with constant $\tfrac1{2\sqrt\delta}$, but on no rectangle that contains a point of the $x$-axis.
  • 5 Give a function, continuous on $\mathbb{R}^2$ and Lipschitz in $y$, for which $\partial f/\partial y$ fails to exist at infinitely many values of $y$.
  • 6 Find the smallest Lipschitz constant of $f(x,y)=xe^{-y^2}$ on $[0,1]\times\mathbb{R}$.
  • 7 Show that $f(x,y)=\sqrt{|x|}+y$ is Lipschitz in $y$ on $\mathbb{R}^2$ although it is not Lipschitz in $x$ near $x=0$. What does this give for the IVP $y'=\sqrt{|x|}+y$, $y(x_0)=y_0$?
  • 8 Show that the integral equation $y(x)=\int_0^x\sqrt{|y(t)|}\,dt$ has infinitely many continuous solutions on $[0,\infty)$.
1 $y(x)=1+\int_0^x\left(t^2+y(t)^2\right)dt$. $\varphi_1=1+\int_0^x(t^2+1)\,dt=1+x+\tfrac{x^3}3$. $\varphi_2=1+\int_0^x\left(t^2+\left(1+t+\tfrac{t^3}3\right)^2\right)dt=1+x+x^2+\tfrac{2x^3}{3}+\tfrac{x^4}{6}+\tfrac{2x^5}{15}+\tfrac{x^7}{63}$.
2 By Theorem 1.1 this is $y'=y^2$, $y(0)=1$, whose maximal solution is $y=\tfrac1{1-x}$ on $(-\infty,1)$ (Example 6). Uniqueness ($f=y^2$ locally Lipschitz) shows it is the only one.
3 (a) $|f_y|=2|xy|\le6$: $L=6$. (b) $|e^{-x}\cos y|\le1$ and the strip is convex in $y$: $L=1$ (attained at $x=0$, $y\to0$). (c) Not Lipschitz: $|h^{2/3}-0|/h=h^{-1/3}\to\infty$. (d) $L=2$, as in Example 2(c).
4 For $|y_1|,|y_2|\ge\delta$: $\left|\sqrt{|y_1|}-\sqrt{|y_2|}\right|=\dfrac{\big||y_1|-|y_2|\big|}{\sqrt{|y_1|}+\sqrt{|y_2|}}\le\dfrac{|y_1-y_2|}{2\sqrt\delta}$ — this works even when $y_1,y_2$ have opposite signs, where the mean value theorem would not apply. Near the axis, $\sqrt{h}/h=h^{-1/2}\to\infty$.
5 $f(x,y)=|\sin y|$: $\big||\sin y_1|-|\sin y_2|\big|\le|\sin y_1-\sin y_2|\le|y_1-y_2|$, but $f$ has corners at $y=k\pi$, $k\in\mathbb{Z}$.
6 $f_y=-2xye^{-y^2}$. $g(y)=2ye^{-y^2}$ has $g'=2(1-2y^2)e^{-y^2}$, so $\max|g|=g\left(\tfrac1{\sqrt2}\right)=\sqrt2e^{-1/2}$. With $0\le x\le1$: $L=\sqrt{2/e}\approx0.858$.
7 $|f(x,y_1)-f(x,y_2)|=|y_1-y_2|$ — the $x$-dependence cancels, because the two points share the same $x$. By Theorem 1.6 the linear IVP has a unique solution on $\mathbb{R}$ for every $(x_0,y_0)$. Regularity in $x$ is irrelevant beyond continuity.
8 By Theorem 1.1 these are the solutions of $y'=\sqrt{|y|}$, $y(0)=0$ on $[0,\infty)$. For each $c\ge0$, $y_c=0$ on $[0,c]$ and $y_c=\tfrac{(x-c)^2}{4}$ for $x\ge c$ works: $y_c'=\tfrac{x-c}2=\sqrt{y_c}$ there, and both one-sided derivatives at $c$ are $0$. Together with $y\equiv0$ these are infinitely many solutions.
II
Picard iterates and intervals of existence.
  • 1 Compute the Picard iterates for $y'=-y$, $y(0)=2$, find $\varphi_n$ and its limit.
  • 2 For $y'=x+y$, $y(0)=1$, compute $\varphi_1$, $\varphi_2$ and compare with the exact solution.
  • 3 For $y'=y^2$, $y(0)=1$, compute $\varphi_1,\varphi_2$ and compare them with the Taylor series of the exact solution.
  • 4 For $y'=y^2$, $y(0)=1$ on $|x|\le1$, $|y-1|\le b$, find $h(b)$ and its largest value.
  • 5 For $y'=x^2+y^2$, $y(0)=0$ on the square $|x|\le a$, $|y|\le a$, find $h$ as a function of $a$ and the best choice of $a$.
  • 6 For $y'=1+\tfrac12\sin y$, $y(0)=0$ on $[0,1]$, use the bound of Theorem 1.6 (with $M_0=\max|f(t,y_0)|$) to find $n$ such that $|\varphi-\varphi_n|\lt10^{-6}$ on $[0,1]$.
  • 7 Show that the Picard iterates of $y'=3y^{2/3}$, $y(0)=0$ converge, and explain why this does not contradict Example 5.
  • 8 Prove that $y'=\sin(xy)$, $y(0)=y_0$ has a unique solution on all of $\mathbb{R}$.
1 $\varphi_1=2-2x$, $\varphi_2=2-2x+x^2$, $\varphi_3=2-2x+x^2-\tfrac{x^3}{3}$; by induction $\varphi_n=2\sum_{k=0}^n\tfrac{(-x)^k}{k!}\to2e^{-x}$ uniformly on bounded intervals.
2 $\varphi_1=1+x+\tfrac{x^2}2$, $\varphi_2=1+x+x^2+\tfrac{x^3}6$. The linear equation gives $y=2e^x-x-1=1+x+x^2+\tfrac{x^3}{3}+\cdots$: $\varphi_2$ agrees up to $x^2$.
3 $\varphi_1=1+x$, $\varphi_2=1+x+x^2+\tfrac{x^3}3$, while $\tfrac1{1-x}=1+x+x^2+x^3+\cdots$. Each iteration fixes (at least) one more Taylor coefficient.
4 $M=(1+b)^2$, $h(b)=\min\left(1,\tfrac{b}{(1+b)^2}\right)=\tfrac{b}{(1+b)^2}$ (always $\le\tfrac14$). Its maximum is at $b=1$: $h=\tfrac14$. The true life span forward is $1$.
5 $M=a^2+a^2=2a^2$, so $h=\min\left(a,\tfrac{a}{2a^2}\right)=\min\left(a,\tfrac1{2a}\right)$, largest when $a=\tfrac1{2a}$: $a=\tfrac1{\sqrt2}$, $h=\tfrac1{\sqrt2}\approx0.707$. (The actual solution exists on $[0,\omega_+)$ with $\omega_+\approx2.003$; see IV.8.)
6 $L=\tfrac12$ (since $|\tfrac12\cos y|\le\tfrac12$), $M_0=\max|f(t,0)|=1$, length $1$. The bound is $\dfrac{M_0}{L}\dfrac{(L\cdot1)^{n+1}}{(n+1)!}e^{L}=2\dfrac{(1/2)^{n+1}}{(n+1)!}e^{1/2}$: about $5.1\times10^{-6}$ for $n=6$ and $3.2\times10^{-7}$ for $n=7$. So $n=7$ suffices.
7 $\varphi_0=0$ and $f(t,0)=0$, so $\varphi_1=0$, and inductively all $\varphi_n=0$: they converge to the solution $y\equiv0$. Picard iteration constructs a fixed point of the integral operator when it converges; without a Lipschitz condition nothing says it is the only one.
8 $\left|\partial_y\sin(xy)\right|=|x\cos(xy)|\le T$ on the strip $[-T,T]\times\mathbb{R}$. Theorem 1.6 gives a unique solution on $[-T,T]$ for every $T$; by uniqueness these agree on overlaps and define one solution on $\mathbb{R}$.
III
Maximal intervals, non-uniqueness and blow-up.
  • 1 Find the maximal interval of the solution of $y'=1+y^2$, $y(0)=1$.
  • 2 Solve $y'=xy^2$, $y(0)=2$, and find the maximal interval.
  • 3 Solve $y'=-y^3$, $y(0)=1$. In which direction does the solution blow up?
  • 4 Find all solutions of $y'=\sqrt{|y|}$, $y(0)=0$ on $\mathbb{R}$.
  • 5 Prove that $y'=1+y^{2/3}$, $y(0)=0$ has exactly one solution near $x=0$.
  • 6 For the logistic equation $y'=y(1-y)$ show that if $0\lt y_0\lt1$ the solution exists on $\mathbb{R}$ and stays in $(0,1)$, and that if $y_0\gt1$ the maximal interval is $\left(\log(1-1/y_0),\infty\right)$.
  • 7 Give a continuous $f$ on $\mathbb{R}^2$ such that the IVP through every point of the $x$-axis has more than one solution.
  • 8 Let $f$ be continuous and locally Lipschitz in $y$ on $\mathbb{R}^2$ and bounded, $|f|\le M$. Prove that every maximal solution of $y'=f(x,y)$ is defined on $\mathbb{R}$.
1 $\arctan y=x+\tfrac\pi4$, $y=\tan\left(x+\tfrac\pi4\right)$, finite while $-\tfrac\pi2\lt x+\tfrac\pi4\lt\tfrac\pi2$: maximal interval $\left(-\tfrac{3\pi}4,\tfrac\pi4\right)$.
2 $-\tfrac1y=\tfrac{x^2}2+C$, $C=-\tfrac12$: $y=\dfrac{2}{1-x^2}$ on $(-1,1)$, blowing up at both ends.
3 $y^{-3}y'=-1$: $-\tfrac1{2y^2}=-x-\tfrac12$, so $y=\dfrac1{\sqrt{1+2x}}$ on $\left(-\tfrac12,\infty\right)$. It decays forward and blows up backward at $x=-\tfrac12$: $-y^3$ pulls towards $0$ for increasing $x$.
4 Solutions are non-decreasing ($y'\ge0$). For $y\lt0$, $y'=\sqrt{-y}$ is solved by $y=-\tfrac{(x-d)^2}{4}$ for $x\le d$ (its derivative $-\tfrac{x-d}2=\tfrac{|x-d|}{2}$ there). All solutions: $$y=\begin{cases}-\tfrac{(x-d)^2}4,&x\lt d,\\0,&d\le x\le c,\\\tfrac{(x-c)^2}4,&x\gt c,\end{cases}\qquad -\infty\le d\le0\le c\le\infty,$$ by the argument of Example 5.
5 Any solution has $y'\ge1$, so it is strictly increasing. Put $G(y)=\int_0^y\frac{ds}{1+s^{2/3}}$, a $C^1$ strictly increasing bijection of $\mathbb{R}$ onto its range. Then $\tfrac{d}{dx}G(y(x))=\tfrac{y'}{1+y^{2/3}}=1$ and $G(y(0))=0$, so $G(y(x))=x$, $y=G^{-1}(x)$: every solution equals this one.
6 $y\equiv0$ and $y\equiv1$ are solutions and $f$ is $C^1$, so no other solution can touch them: a solution starting in $(0,1)$ stays there, hence is bounded, and by Theorem 1.5 exists on $\mathbb{R}$. Explicitly $y=\dfrac{1}{1+\frac{1-y_0}{y_0}e^{-x}}$; for $y_0\gt1$ the denominator vanishes when $e^{-x}=\tfrac{y_0}{y_0-1}$, i.e. at $x=\log\left(1-\tfrac1{y_0}\right)\lt0$, where $y\to+\infty$. Forward, $y\downarrow1$.
7 $f(x,y)=3|y|^{2/3}$ (the right side of Example 5): through $(c,0)$ pass $y\equiv0$ and $y=(x-c)^3$.
8 If $\omega_+\lt\infty$, then for $x_0\le x\lt\omega_+$, $|\varphi(x)-y_0|\le M(x-x_0)\le M(\omega_+-x_0)$, so the graph stays in the compact set $[x_0,\omega_+]\times[y_0-N,y_0+N]$ with $N=M(\omega_+-x_0)$, contradicting Theorem 1.5. Similarly $\omega_-=-\infty$.
IV
Gronwall, dependence, and solving tools.
  • 1 Let $u\ge0$ be continuous on $[x_0,x_1]$ with $u(x)\le L\int_{x_0}^xu(t)\,dt$. Prove from Theorem 1.7 that $u\equiv0$, and deduce uniqueness for Lipschitz $f$ in two lines.
  • 2 Let $y'=\cos y$, $y(0)=0$ and $z'=\cos z+0.01$, $z(0)=0$. Bound $|y(x)-z(x)|$ on $[0,1]$.
  • 3 Solve $(2xy+3x^2)\,dx+(x^2+2y)\,dy=0$ with $y(0)=1$.
  • 4 Find an integrating factor for $(x^2+y^2+x)\,dx+xy\,dy=0$ and solve.
  • 5 Solve $y'-\dfrac{2y}{x}=-x^2y^2$, $y(1)=1$, and find the maximal interval.
  • 6 Solve $y'+y\tan x=\sec x$, $y(0)=1$. The formula you obtain is defined on $\mathbb{R}$; what is the maximal interval of the solution of this IVP, and why?
  • 7 Prove: if $f$ is continuous on $D$ and non-increasing in $y$ for each fixed $x$, then the IVP has at most one solution on $[x_0,x_0+\delta)$. Apply this to $y'=-y^{1/3}$, $y(0)=0$, and show that uniqueness fails backwards.
  • 8 Let $y$ be the maximal solution of $y'=x^2+y^2$, $y(0)=0$, on $[0,\omega_+)$. Show that $y(x)\ge\tfrac{x^3}{3}$, and that $\omega_+\le1+\tfrac\pi2-\arctan\tfrac13\approx2.249$.
1 Theorem 1.7 with $C=0$, $k=L$: $u\le0\cdot e^{L(x-x_0)}=0$, and $u\ge0$. If $y,z$ solve the same IVP, $u=|y-z|\le\int_{x_0}^x|f(t,y)-f(t,z)|\le L\int_{x_0}^xu$, so $u\equiv0$.
2 $L=1$ ($|\partial_y\cos y|\le1$), $\varepsilon=0.01$, $\delta=0$: $|y-z|\le0.01(e^x-1)\le0.01(e-1)\approx0.0172$ on $[0,1]$.
3 $M_y=2x=N_x$: exact. $\Phi=x^2y+x^3+y^2$ ($\Phi_x=2xy+3x^2$, $\Phi_y=x^2+2y$). $\Phi(0,1)=1$: the solution satisfies $x^2y+x^3+y^2=1$; solving the quadratic in $y$ with $y(0)=1$: $y=\tfrac12\left(-x^2+\sqrt{x^4-4x^3+4}\right)$.
4 $M_y-N_x=2y-y=y$, and $\tfrac{y}{N}=\tfrac1x$: $\mu=x$. Then $(x^3+xy^2+x^2)\,dx+x^2y\,dy=0$ is exact with $\Phi=\tfrac{x^4}4+\tfrac{x^2y^2}2+\tfrac{x^3}3$; solutions $3x^4+6x^2y^2+4x^3=C$.
5 Bernoulli, $n=2$: $v=\tfrac1y$ gives $v'+\tfrac{2v}x=x^2$, $(x^2v)'=x^4$, $x^2v=\tfrac{x^5}{5}+C$, $C=\tfrac45$: $$y=\frac{5x^2}{x^5+4}.$$ The equation is defined for $x\gt0$ (because of $\tfrac{2y}{x}$), and the formula is finite there, so the maximal interval is $(0,\infty)$.
6 $\mu=e^{\int\tan x}=\sec x$: $(y\sec x)'=\sec^2x$, $y\sec x=\tan x+C$, $y=\sin x+C\cos x$, $C=1$: $y=\sin x+\cos x$. The equation itself (its coefficients $\tan x$, $\sec x$) is defined only on intervals avoiding odd multiples of $\tfrac\pi2$; the IVP lives on $\left(-\tfrac\pi2,\tfrac\pi2\right)$, and that is its maximal interval even though the formula extends.
7 If $y,z$ are solutions, $\tfrac{d}{dx}(y-z)^2=2(y-z)\big(f(x,y)-f(x,z)\big)\le0$, because $f(x,\cdot)$ non-increasing makes the two factors of opposite signs. So $(y-z)^2$ is non-increasing and starts at $0$: $y=z$ for $x\ge x_0$. $f=-y^{1/3}$ is non-increasing: forward the only solution is $y\equiv0$. Backwards, $y=\pm\left(\tfrac{2(c-x)}3\right)^{3/2}$ for $x\le c$ ($c\le0$), $0$ for $x\ge c$, are further solutions.
8 $y'\ge x^2$, so $y(x)\ge\tfrac{x^3}3$; in particular $y(1)\ge\tfrac13$ (numerically $y(1)\approx0.350$). For $x\ge1$, $y'\ge1+y^2$, so $\tfrac{d}{dx}\arctan y=\tfrac{y'}{1+y^2}\ge1$ and $\arctan y(x)\ge\arctan y(1)+x-1$. Since $\arctan\lt\tfrac\pi2$, $x-1\lt\tfrac\pi2-\arctan y(1)\le\tfrac\pi2-\arctan\tfrac13$ for every $x\lt\omega_+$. (Numerically $\omega_+\approx2.0031$.)
🎯 Exam corner CSIR Part BCSIR Part CGATE MA

E1. CSIR Part B GATE MA For the IVP $y'=1+y^2$, $y(0)=0$, considered on the rectangle $R=\{|x|\le2,\ |y|\le3\}$, the Picard–Lindelöf theorem guarantees a unique solution on $|x|\le h$ with $h=$ (A) $2$; (B) $\tfrac{3}{10}$; (C) $\tfrac12$; (D) $\tfrac\pi2$.

Idea $h=\min(a,b/M)$ with $M=\max_R|f|$ on the given rectangle.
Compute $a=2$, $b=3$, $M=1+9=10$: $h=\min(2,0.3)=0.3$. (C) is the best value over all rectangles (Example 4), and (D) is the true half-life of $\tan x$ — neither is what the theorem gives on this $R$.
Answer (B).

E2. CSIR Part B The maximal interval of existence of the solution of $y'=e^{y}$, $y(0)=0$ is (A) $\mathbb{R}$; (B) $(-\infty,1)$; (C) $(-1,1)$; (D) $(-\infty,e)$.

Solve $e^{-y}y'=1$: $-e^{-y}=x-1$, so $y=-\log(1-x)$.
Interval Defined for $x\lt1$, and $y\to+\infty$ as $x\to1^-$; backward it exists for all $x$ ($y\to-\infty$ only as $x\to-\infty$).
Answer (B).

E3. CSIR Part B Which of the following functions satisfies a Lipschitz condition in $y$ on the whole plane $\mathbb{R}^2$? (A) $f(x,y)=xy$; (B) $f(x,y)=y^{1/3}$; (C) $f(x,y)=e^{-|y|}\cos x$; (D) $f(x,y)=y^2$.

Options (A) $|xy_1-xy_2|=|x||y_1-y_2|$: the constant would have to exceed every $|x|$ — no. (B) fails at $y=0$. (D) $|y_1+y_2|$ unbounded. (C) $\left|e^{-|y_1|}-e^{-|y_2|}\right|\le\big||y_1|-|y_2|\big|\le|y_1-y_2|$ (the map $s\mapsto e^{-s}$ has slope at most $1$ on $[0,\infty)$) and $|\cos x|\le1$: $L=1$.
Answer (C).

E4. CSIR Part C For $\alpha\gt0$ consider $y'=|y|^{\alpha}$, $y(0)=0$. Which of the following statements are true? (One or more options may be correct.) (A) For every $\alpha\ge1$ the IVP has exactly one solution on $\mathbb{R}$. (B) For every $0\lt\alpha\lt1$ the IVP has infinitely many solutions on $\mathbb{R}$. (C) For $\alpha=2$, the solution with $y(0)=1$ instead of $y(0)=0$ is not defined on all of $[0,\infty)$. (D) For $\alpha=\tfrac12$ every solution vanishes on $(-\infty,0]$.

(A) True For $\alpha=1$, $|y|$ is Lipschitz; for $\alpha\gt1$, $|y|^\alpha$ is $C^1$ (derivative $\alpha|y|^{\alpha-1}\operatorname{sgn}y$). So $f$ is locally Lipschitz, $y\equiv0$ is a global solution, and by Theorem 1.5 it is the only one.
(B) True For $x\ge c\ge0$, $y=\big((1-\alpha)(x-c)\big)^{1/(1-\alpha)}$ satisfies $y'=y^\alpha$ and has zero derivative at $c$; glued to $0$ on $(-\infty,c]$ it solves the IVP for every $c\ge0$.
(C) True $y'=y^2$, $y(0)=1$ gives $y=\tfrac1{1-x}$, which blows up at $x=1$.
(D) False $y=-\tfrac{(x-d)^2}4$ for $x\le d$, $y=0$ for $x\ge d$, with any $d\lt0$, is a solution that is nonzero on $(-\infty,d)$ (Exercise III.4).
Answer (A), (B), (C).

E5. CSIR Part C GATE MA Let $f:\mathbb{R}^2\to\mathbb{R}$ be continuous with $|f(x,y_1)-f(x,y_2)|\le|y_1-y_2|$ for all $x,y_1,y_2$. Which of the following are true? (A) Every IVP $y'=f(x,y)$, $y(x_0)=y_0$ has a unique solution defined on $\mathbb{R}$. (B) The graphs of two different solutions may intersect. (C) If $y,z$ are solutions with $y(0)=0$ and $z(0)=10^{-3}$, then $|y(1)-z(1)|\le e\cdot10^{-3}$. (D) Every solution is bounded on $\mathbb{R}$.

(A) True Theorem 1.6 on every strip $[-T,T]\times\mathbb{R}$, then uniqueness to glue.
(B) False At an intersection point both would solve the same IVP; by uniqueness they would coincide.
(C) True Theorem 1.8 with $L=1$, $\varepsilon=0$: $|y(1)-z(1)|\le10^{-3}e^{1}$.
(D) False $f(x,y)=y$ satisfies the hypothesis and has the solution $e^x$.
Answer (A) and (C).

E6. CSIR Part C Let $\varphi_n$ be the Picard iterates of $y'=y^2$, $y(0)=1$ (with $\varphi_0=1$). Which of the following are true? (A) $\varphi_2(x)=1+x+x^2+\tfrac{x^3}{3}$. (B) $\varphi_n\to\tfrac{1}{1-x}$ uniformly on $\left[-\tfrac14,\tfrac14\right]$. (C) The solution of the IVP can be extended to all of $\mathbb{R}$. (D) The $\varphi_n$ converge uniformly on $[0,2]$.

(A) True $\varphi_1=1+x$, $\varphi_2=1+\int_0^x(1+t)^2dt=1+x+x^2+\tfrac{x^3}3$.
(B) True On $|x|\le1$, $|y-1|\le1$: $M=4$, $h=\min(1,\tfrac14)=\tfrac14$; Theorem 1.3 gives uniform convergence on $\left[-\tfrac14,\tfrac14\right]$ to the solution $\tfrac1{1-x}$.
(C) False The maximal interval is $(-\infty,1)$ (Example 6).
(D) False If $\varphi_n\to\varphi$ uniformly on $[0,2]$, the $\varphi_n$ would be uniformly bounded and, passing to the limit in $\varphi_{n+1}=1+\int_0^x\varphi_n^2$, $\varphi$ would be a continuous solution on $[0,2]$ — impossible, since the solution blows up at $1$. (Numerically $\varphi_6(1.5)\approx8.7\times10^3$.)
Answer (A) and (B).
🧑‍🏫 For the teacher

Suggested plan — 3 lectures × 60 min.

  1. The question and the tools. Open with the three pathologies ($y'=3y^{2/3}$, $y'=y^2$, the discontinuous $f$ of pitfall 2) before any theorem. Theorem 1.1 with Example 1; Definition 1.2 and Theorem 1.2 with Example 2 worked on the board. Project Interactive 2 in its first mode while students try to "decide" the solution through $(0,0)$.
  2. Picard–Lindelöf. Example 3 first (students compute iterates and recognise $e^{x^2}-1$), then the full proof of Theorem 1.3, emphasising where $h\le b/M$ and where $L$ are used. Interactive 1: modes 1–3, then mode 4 beyond $x=1$. Example 4, Theorem 1.4 (sketch only).
  3. Maximal solutions, Gronwall, tools. Example 5 (gluing), Theorem 1.5 with the picture of a solution leaving a compact set, Example 6 with Interactive 2 in its second mode, Theorem 1.6 (linear equations never blow up). Gronwall and Theorem 1.8; Example 7; finish with Exam-corner E1 and E4 live.

Discussion questions. (1) Picard iterates for $y'=3y^{2/3}$, $y(0)=0$ converge — so why is the solution not unique? What exactly does the Lipschitz condition add? (2) Is there an equation $y'=f(y)$ with $f$ continuous whose solutions are unique although $f$ is nowhere Lipschitz near some point? (3) A computer integrates $y'=y^2$, $y(0)=1$ to $x=2$ without complaint and prints a finite number. What has gone wrong, and how would Theorem 1.5 warn you?

Misconceptions to watch for. "Continuous $f$ ⇒ unique solution"; "$f$ differentiable everywhere ⇒ solution exists for all $x$"; confusing Lipschitz in $y$ with Lipschitz in $(x,y)$; reading $h=\min(a,b/M)$ as the maximal interval; thinking that the explicit formula's domain is the maximal interval (Example 7, Exercise IV.6); forgetting $k\ge0$ in Gronwall.

Suggested homework. I(1, 3, 4, 7), II(2, 4, 6, 7), III(1, 4, 6, 8), IV(2, 5, 7, 8); Exam corner E2, E5, E6.

🔬 Research–Teaching NexusResearch-tutored · optional, about 60 minutes

Research question. When does a model "explode" in finite time — and can we trap the explosion time between two numbers we can prove, before any computer is switched on?

Why it matters. Finite-time blow-up is not only a mathematical curiosity. Thermal runaway in batteries and chemical reactors, the ignition of a fuel mixture, and the collapse of a self-gravitating cloud are all modelled by equations whose solutions blow up; engineers in battery safety and combustion research need the blow-up time (the "ignition delay") and its sensitivity to the data. The Riccati equation $y'=x^2+y^2$ is the simplest model with a source that grows in time and a superlinear feedback.

What to do.

  1. Prove a lower bound for $\omega_+$ for $y'=x^2+y^2$, $y(0)=0$: on $[0,1]$ compare with $y'\le1+y^2$ (show $\arctan y(x)\le x$, so the solution survives past $x=1$); then improve it on $[1,\omega_+)$ using $y'\le x_1^2+y^2$ on $[1,x_1]$.
  2. Prove an upper bound as in Exercise IV.8, and sharpen it by using a better lower estimate for $y(1)$ (for example $y\ge\varphi_2$, a Picard iterate — justify why!).
  3. Integrate numerically with a stepsize-controlled method until $y$ exceeds $10^8$, for several tolerances, and estimate $\omega_+$ to five digits.
  4. Repeat for $y'=x^2+y^2$, $y(0)=y_0$ with $y_0\in\{-1,0,1,2\}$ and plot $\omega_+(y_0)$. Compare with Theorem 1.8: is $\omega_+$ a continuous function of $y_0$?

Tools: Pen and paper for the bounds; Python with scipy.integrate.solve_ivp (with an event that stops the integration at a large value) or a spreadsheet RK4 with a halving step.

Hand in: A one-page note with your proved bracket for $\omega_+$, the numerical estimate, and the graph of $\omega_+(y_0)$.

Going further. The substitution $y=-u'/u$ turns this Riccati equation into the linear equation $u''+x^2u=0$; blow-up of $y$ corresponds to the first positive zero of $u$. Read about this link and explain why Sturm's comparison theorem (§2.3) gives bounds on $\omega_+$.

Instructor note: for $y(0)=0$, $\omega_+\approx2.00315$ and $y(1)\approx0.3502$. The crude bracket from Exercise IV.8 is $(1,\,2.249)$; using $y(1)\ge\varphi_2(1)=\tfrac13+\tfrac1{63}=\tfrac{22}{63}$ improves the upper bound to about $2.235$. The blow-up time decreases as $y_0$ increases, continuously, as students should verify.

🧪 Self-check quiz

0/5 correct

Q1. Which hypotheses on $f$ in a rectangle around $(x_0,y_0)$ guarantee that $y'=f(x,y)$, $y(x_0)=y_0$ has exactly one local solution?

Q2. How many solutions does $y'=y^{1/3}$, $y(0)=0$ have on $[0,\infty)$?

Q3. The maximal interval of the solution of $y'=y^2$, $y(0)=-1$ is

Q4. For $y'=y$, $y(0)=1$ on the rectangle $|x|\le1$, $|y-1|\le1$, Theorem 1.3 guarantees a solution on $|x|\le h$ with $h=$

Q5. If $u$ is continuous on $[0,1]$ and $u(x)\le2+\int_0^x3u(t)\,dt$, Gronwall's inequality gives