Picard-Lindelof existence and uniqueness for ODEs

Intuition Beginner

A differential equation tells you the slope of a curve at every point. Add one starting value, and you face a sharp question: is there a curve that obeys those slopes and passes through that start — and is it the only one? This is the question of existence and uniqueness. Most physical models assume the answer is yes: a pendulum released from a known angle with a known speed has exactly one future. But the answer is not automatic. Some slope rules are so badly behaved that many curves fit, or none do. Picard and Lindelof isolated the clean condition that settles the matter.

The condition is a steady-steepness rule: if you nudge the starting value a little, the slope nudges by a proportional amount, never more. Under that rule, a process called Picard iteration builds the answer. You begin with a flat guess (the starting value held constant), feed it into the slope rule to get a better guess, and feed that guess back in. Each pass refines the curve. The guesses settle down to one limit, and that limit is the unique solution. Picture a navigator who keeps correcting a heading against the local compass field until the traced path stops moving.

Why care? Because before solving anything, you want assurance the problem is well-posed: one and only one future flows from each starting state. Without that, a computer simulation could drift between equally valid answers, and predictions would mean nothing. This unit delivers that assurance and names the precise price: continuity of the slope rule buys existence; a Lipschitz bound on steepness buys uniqueness. Drop the bound and existence survives (Peano) but uniqueness can fail — two curves can leave the same start and diverge.

Visual Beginner

The picture shows two panels side by side. The left panel plots the first four Picard iterates for $y^{\prime }=y$, $y(0)=1$: a flat line ${\varphi }_0=1$, then a slanted line ${\varphi }_1=1+x$, then ${\varphi }_2=1+x+x^2/2$, then ${\varphi }_3=1+x+x^2/2+x^3/6$, each one hugging the exponential $e^x$ more tightly. The right panel shows the Peano failure: the equation $y^{\prime }=\sqrt{|y|}$, $y(0)=0$, carries not one but a whole family of solution curves leaving the origin — a flat curve $y=0$ together with parabolas $y=x^2/4$ that branch off at every starting delay.

The contrast is the whole story in one image: a Lipschitz bound (left) funnels the iteration onto a single curve; continuity alone (right) leaves the future underdetermined.

Worked example Beginner

Run Picard iteration for $y^{\prime }=2y$ with $y(0)=1$, and watch the guesses close in on the true curve at $x=0.5$.

Step 0. Start with the flat guess ${\varphi }_0(x)=1$ — the starting value held constant.

Step 1. The slope rule is "slope $=2y$". On ${\varphi }_0$ the slope is $2\cdot 1=2$ everywhere. Accumulate this slope from $0$ to $x$: the improved guess is ${\varphi }_1(x)=1+2x$.

Step 2. On ${\varphi }_1$ the slope is $2(1+2t)=2+4t$. Accumulate from $0$ to $x$: you add $2x+2x^2$, giving ${\varphi }_2(x)=1+2x+2x^2$.

Step 3. On ${\varphi }_2$ the slope is $2+4t+4t^2$. Accumulate to add $2x+2x^2+\frac{4}{3}{x}^3$, giving ${\varphi }_3(x)=1+2x+2x^2+\frac{4}{3}{x}^3$.

The exact solution is $y=e^{2x}=1+2x+2x^2+\frac{4}{3}{x}^3+\cdots$. At $x=0.5$ the guesses read ${\varphi }_0=1$, ${\varphi }_1=2$, ${\varphi }_2=2.5$, ${\varphi }_3\approx 2.667$, marching toward $e^1\approx 2.718$.

What this tells us: each pass layers one more power of $x$ onto the guess, exactly rebuilding the exponential one term at a time. The iteration cannot help but converge, because the slope rule's steepness is bounded.

Check your understanding Beginner

Formal definition Intermediate+

Definition (Initial value problem). Let $D\subset \mathbb{R}\times {\mathbb{R}}^n$ be open and let $f:D\to {\mathbb{R}}^n$ be continuous. The initial value problem (IVP) is

$$y^{\prime }=f(x,y),y(x_0)=y_0,(x_0,y_0)\in D.$$

A solution on an interval $I\ni x_0$ is a $C^1$ map $y:I\to {\mathbb{R}}^n$ with $(x,y(x))\in D$ satisfying the ODE and the initial condition. The vector-valued formulation covers a single higher-order equation after the standard reduction 02.08.02, so the theory below applies to every order at once.

Definition (Lipschitz in $y$). The function $f$ is Lipschitz in $y$ on $D$ if there exists $L\ge 0$ with

$$\Vert f(x,y_1)-f(x,y_2)\Vert \le L\Vert y_1-y_2\Vert$$

for all $(x,y_1),(x,y_2)\in D$. It is locally Lipschitz in $y$ if every point of $D$ has a neighbourhood on which such a bound holds. A sufficient condition, obtained from the mean value theorem 02.05.02, is that $\partial f/\partial y$ exists and is bounded on $D$.

Definition (Picard operator). Setting the IVP in integral form via the fundamental theorem of calculus 02.04.04,

$$y(x)=y_0+{\int }_{x_0}^{x}f(t,y(t))dt,$$

defines the Picard operator $T$ on continuous curves by $(T\phi )(x)=y_0+{\int }_{x_0}^{x}f(t,\phi (t))dt$. A solution of the IVP is precisely a fixed point of $T$.

Counterexamples to common slips

• Continuity is weaker than Lipschitz. The map $f(y)=\sqrt{|y|}$ is continuous on $\mathbb{R}$ but not Lipschitz near $0$: the difference quotient $|\sqrt{|y_1|}-\sqrt{|y_2|}|/|y_1-y_2|$ blows up as $y_1,y_2\to 0$. The IVP $y^{\prime }=\sqrt{|y|}$, $y(0)=0$, has the constant solution $y\equiv 0$ and also $y(x)=x^2/4$ (and the delayed parabolas $y(x)=\frac{1}{4}\max (0,x-c{)}^2$ for any $c\ge 0$).

• A finite Lipschitz constant is local, not global. The equation $y^{\prime }=y^2$ has $f(y)=y^2$, which is locally Lipschitz everywhere but not globally Lipschitz on $\mathbb{R}$. Existence and uniqueness hold only up to a finite blow-up time; the solution $y=1/(1-x)$ from $y(0)=1$ escapes at $x=1$.

• Existence on a strip is not existence for all time. The Picard-Lindelof theorem is local: it produces a solution on $|x-x_0|\le h$ with $h=\min (a,b/M)$. Extending to a maximal interval is a separate argument (the blow-up alternative below).

Key theorem with proof Intermediate+

Theorem (Picard-Lindelof: local existence and uniqueness). Let $f$ be continuous on the rectangle $R=\{(x,y):|x-x_0|\le a,\Vert y-y_0\Vert \le b\}$ and Lipschitz in $y$ on $R$ with constant $L$. Set $M={\sup }_R\Vert f\Vert$ and $h=\min (a,b/M)$ (with $h=a$ if $M=0$). Then the IVP $y^{\prime }=f(x,y)$, $y(x_0)=y_0$, has a unique solution on $[x_0-h,x_0+h]$.

The proof rests on a single abstract fact, which we state as a lemma so the contraction-mapping theorem is visible in the argument 02.11.04.

Lemma (Banach contraction mapping principle). Let $(X,d)$ be a non-empty complete metric space and $T:X\to X$ a contraction, i.e. $d(Tu,Tv)\le qd(u,v)$ for some $q<1$. Then $T$ has a unique fixed point, and the iterates $T^n{u}_0$ converge to it for every $u_0\in X$.

Proof of the theorem. By the fundamental theorem of calculus 02.04.04, a $C^1$ curve $y$ solves the IVP on an interval $I\ni x_0$ if and only if it satisfies the integral equation $y=Ty$ with $T$ the Picard operator above. We exhibit a complete metric space on which $T$ is a contraction.

Let $I=[x_0-h,x_0+h]$ and let $X=\{\phi \in C(I,{\mathbb{R}}^n):\Vert \phi -y_0{\Vert }_{\infty }\le b\}$, a closed ball in the Banach space $C(I,{\mathbb{R}}^n)$ with the supremum norm 02.11.04; hence $X$ is complete.

$T$ maps $X$ into itself. For $\phi \in X$ and $x\in I$,

$$\Vert (T\phi )(x)-y_0\Vert =\Vert {\int }_{x_0}^{x}f(t,\phi (t))dt\Vert \le M|x-x_0|\le Mh\le b,$$

so $T\phi \in X$.

$T$ is a contraction in the Bielecki norm $\Vert \phi {\Vert }_{\lambda }={\sup }_{x\in I}{e}^{-\lambda |x-x_0|}\Vert \phi (x)\Vert$ with $\lambda >L$. This norm is equivalent to the supremum norm, so completeness is preserved. For $\phi ,\psi \in X$,

$$\begin{array}{rl}\Vert (T\phi )(x)-(T\psi )(x)\Vert & \le {\int }_{x_0}^{x}L\Vert \phi (t)-\psi (t)\Vert dt\\ & ={\int }_{x_0}^{x}L{e}^{\lambda |t-x_0|}{e}^{-\lambda |t-x_0|}\Vert \phi (t)-\psi (t)\Vert dt\\ & \le L\Vert \phi -\psi {\Vert }_{\lambda }{\int }_{x_0}^x{e}^{\lambda |t-x_0|}dt\\ & \le \frac{L}{\lambda }{e}^{\lambda |x-x_0|}\Vert \phi -\psi {\Vert }_{\lambda }.\end{array}$$

Multiplying by $e^{-\lambda |x-x_0|}$ and taking the supremum gives $\Vert T\phi -T\psi {\Vert }_{\lambda }\le (L/\lambda )\Vert \phi -\psi {\Vert }_{\lambda }$. Since $L/\lambda <1$, $T$ is a contraction. By the lemma, $T$ has a unique fixed point in $X$, and the Picard iterates ${\phi }_{n+1}=T{\phi }_n$ (with ${\phi }_0\equiv y_0$) converge to it. $\square$

The fixed point is the unique solution that stays in the ball $X$. Any solution on a subinterval must stay in $X$ (the same estimate $\Vert y(x)-y_0\Vert \le M|x-x_0|\le b$ applies), so uniqueness holds on all of $I$.

Bridge. This theorem is the foundational reason initial value problems are well-posed under a Lipschitz condition, and this is exactly the bridge from pointwise regularity of the vector field to a globally meaningful solution operator. The central insight is that the Picard operator is a contraction in a weighted norm, so completeness of the function space 02.11.04 alone forces a unique fixed point. The result builds toward the phase-flow picture 02.12.01 where existence and uniqueness promote the solution into a one-parameter group acting on state space, and appears again in 02.12.08 where Gronwall-type estimates convert the very same Lipschitz data into Lyapunov stability of nearby trajectories. Putting these together, the bridge is that one quantitative bound on how fiercely the field can shear neighbouring points is simultaneously what drives the iteration to converge and what bars nearby solutions from tearing apart.

Exercises Intermediate+

Advanced results Master

Theorem (Peano existence). If $f$ is continuous on the rectangle $R$ (with no Lipschitz hypothesis), then the IVP $y^{\prime }=f(x,y)$, $y(x_0)=y_0$, has at least one solution on $[x_0-h,x_0+h]$ with $h=\min (a,b/M)$. The proof is non-constructive: one builds a sequence of polygonal Euler approximations $y_n$ with step $\to 0$, each uniformly bounded by $M$ and equicontinuous (their slopes are bounded by $M$); the Arzela-Ascoli theorem supplies a uniformly convergent subsequence, and the limit satisfies the integral equation by continuity of $f$. Existence without a uniqueness mechanism is the cost — the Peano example $y^{\prime }=\sqrt{|y|}$ above shows the defect is real, not an artefact of the proof ^[Peano1890].

Theorem (Gronwall's inequality). Integral form. If $\beta \ge 0$ is continuous, $\alpha$ is non-decreasing, and $u(t)\le \alpha (t)+{\int }_{t_0}^t\beta (s)u(s)ds$ on $[t_0,t_1]$, then $u(t)\le \alpha (t)\exp ({\int }_{t_0}^t\beta (s)ds)$. Differential form. If $u$ is differentiable with $u^{\prime }(t)\le \beta (t)u(t)$, then $u(t)\le u(t_0)\exp ({\int }_{t_0}^t\beta (s)ds)$. The estimate is the workhorse of ODE theory: it converts the integral inequality satisfied by the difference of two solutions into an exponential bound forcing that difference to vanish ^{[Gronwall1919]}.

Theorem (maximal interval and the blow-up alternative). If $f$ is continuous and locally Lipschitz in $y$ on $D$, then the solution through $(x_0,y_0)$ extends uniquely to a maximal interval of existence $(a,b)\ni x_0$, and this interval is open. The escape criterion states: if $b<\infty$, then the solution leaves every compact subset of $D$ as $x\to b^{-}$; in the common case $D=\mathbb{R}\times {\mathbb{R}}^n$ this reads $\Vert y(x)\Vert \to \infty$. Contrapositively, a solution that stays bounded on a finite time interval can be extended further. This is why $y^{\prime }=y^2$ escapes at $x=1$ while $y^{\prime }=-y^3$ is global: the former blows up, the latter decays.

Theorem (continuous dependence on initial data). Under the hypotheses of Picard-Lindelof, the solution depends Lipschitz-continuously on the initial value. Concretely, if $y(\cdot ;y_0)$ and $y(\cdot ;z_0)$ are two solutions with the same Lipschitz constant $L$ on a common interval $[x_0,x_0+\tau ]$, then

$$\Vert y(x;y_0)-y(x;z_0)\Vert \le \Vert y_0-z_0\Vert e^{L(x-x_0)},x\in [x_0,x_0+\tau ].$$

The proof applies Gronwall to $\Vert y(x;y_0)-y(x;z_0)\Vert \le \Vert y_0-z_0\Vert +L{\int }_{x_0}^x\Vert y(t;y_0)-y(t;z_0)\Vert dt$.

Theorem (continuous dependence on parameters). If $f=f(x,y,\lambda )$ is continuous and Lipschitz in $y$ (uniformly in $\lambda$) and Lipschitz in $\lambda$, the solution $y(x;y_0,\lambda )$ is continuous — indeed locally Lipschitz — in $(y_0,\lambda )$. This is the theorem that licenses perturbation analysis and numerical simulation: small changes in the model or the data produce small changes in the trajectory.

Theorem (Osgood's uniqueness criterion). Lipschitz is sufficient but not necessary for uniqueness. If $\Vert f(x,y_1)-f(x,y_2)\Vert \le \omega (\Vert y_1-y_2\Vert )$ for a modulus $\omega$ with $\omega (0)=0$ and the Osgood condition ${\int }_{0^{+}}ds/\omega (s)=\infty$, then the IVP has a unique solution. The Lipschitz case is $\omega (r)=Lr$; the borderline $\omega (r)=Lr\log (1/r)$ still gives uniqueness, while $\omega (r)=\sqrt{r}$ (the Peano scale) does not. This sharpens the existence/uniqueness dichotomy to its precise threshold.

Synthesis. The foundational reason the Picard-Lindelof machinery works is that solving an IVP is secretly a fixed-point problem in a complete function space, and this is exactly the bridge between real analysis (continuity, Lipschitz bounds, the mean value theorem 02.05.02) and the geometric theory of flows. The central insight is that a single quantitative estimate — a Lipschitz bound on the vector field — simultaneously yields existence (the iteration converges), uniqueness (Gronwall forces two solutions to coincide), and continuous dependence (the solution map is Lipschitz in the data). Putting these together, the bridge is that well-posedness is one phenomenon with three faces: existence, uniqueness, and stability. The pattern generalises to abstract evolution equations in Banach spaces 02.11.04 and appears again in the phase flow 02.12.01 and in Lyapunov stability 02.12.08, where Gronwall's estimate is the seed of every stability argument. The result builds toward PDE theory, where the same fixed-point reasoning underlies mild solutions, and the Peano-Osgood dichotomy recurs wherever a vector field is continuous but not Lipschitz.

Full proof set Master

Proposition 1 (Gronwall's inequality, integral form). Let $u$ be continuous and non-negative on $[t_0,t_1]$, $\beta \ge 0$ continuous, $\alpha$ non-decreasing and continuous. If $u(t)\le \alpha (t)+{\int }_{t_0}^t\beta (s)u(s)ds$, then $u(t)\le \alpha (t)\exp ({\int }_{t_0}^t\beta (s)ds)$.

Proof. Set $v(t)=\alpha (t)+{\int }_{t_0}^t\beta (s)u(s)ds$, so $u\le v$ and $v(t_0)=\alpha (t_0)$. Because $\alpha$ is non-decreasing and $u\le v$,

$$v^{\prime }(t)={\alpha }^{\prime }(t)+\beta (t)u(t)\le 0+\beta (t)v(t)$$

(in the distributional sense; replace ${\alpha }^{\prime }$ by its measure when $\alpha$ is merely non-decreasing). Multiply by the integrating factor $\mu (t)=\exp (-{\int }_{t_0}^t\beta (s)ds)>0$: $(\mu v{)}^{\prime }=\mu (v^{\prime }-\beta v)\le 0$, so $\mu (t)v(t)\le \mu (t_0)v(t_0)=\alpha (t_0)$. Hence $v(t)\le \alpha (t_0)\exp \left({\int }_{t_0}^t\beta \right)\le \alpha (t)\exp \left({\int }_{t_0}^t\beta \right)$, using that $\alpha$ is non-decreasing. Since $u\le v$, the claim follows. $\square$

Proposition 2 (blow-up alternative). Let $f$ be continuous and locally Lipschitz in $y$ on $D=\mathbb{R}\times {\mathbb{R}}^n$, and let $y:(a,b)\to {\mathbb{R}}^n$ be the maximal solution with $b<\infty$. Then $\Vert y(x)\Vert \to \infty$ as $x\to b^{-}$.

Proof. Suppose for contradiction that $\Vert y(x)\Vert \le K$ on $[x_0,b)$ for some $K$. On the compact rectangle $R=[x_0,b]\times \overline{B(0,K+1)}$ the hypotheses of Picard-Lindelof hold: $f$ is bounded by some $M$ and Lipschitz in $y$ with some constant $L$. Apply the local theorem at the point $(b-\epsilon ,y(b-\epsilon ))$ for $\epsilon >0$ small: it returns a solution on an interval of half-width $h=\min ({\epsilon }^{\prime },1/M)$ that depends only on $R$, hence is bounded below by a constant $\delta >0$ independent of $\epsilon$. Choosing $\epsilon <\delta$ extends the solution past $b$, contradicting maximality of $b$. Therefore no such $K$ exists, i.e. $\Vert y(x)\Vert \to \infty$. $\square$

Proposition 3 (Lipschitz dependence on initial data). Under the hypotheses of the continuous-dependence theorem, $\Vert y(x;y_0)-y(x;z_0)\Vert \le \Vert y_0-z_0\Vert e^{L(x-x_0)}$ for $x\ge x_0$ on the common interval of existence.

Proof. Subtract the integral equations for the two solutions:

$$\Vert y(x;y_0)-y(x;z_0)\Vert \le \Vert y_0-z_0\Vert +{\int }_{x_0}^x\Vert f(t,y(t;y_0))-f(t,y(t;z_0))\Vert dt\le \Vert y_0-z_0\Vert +L{\int }_{x_0}^x\Vert y(t;y_0)-y(t;z_0)\Vert dt.$$

This is a Gronwall inequality with $\alpha =\Vert y_0-z_0\Vert$ and $\beta =L$, so Proposition 1 gives the displayed exponential bound. $\square$

Connections Master

• Banach space fundamentals 02.11.04. The Picard operator acts on the space $C(I,{\mathbb{R}}^n)$ of continuous curves, and the proof needs this space to be complete in the (equivalent) Bielecki norm so that the contraction mapping principle returns a fixed point. Completeness of the ambient function space is the single structural fact the existence argument consumes; without it the iterates would converge to nothing.

• Mean value theorem 02.05.02. The Lipschitz-in-$y$ hypothesis is most often checked via the mean value theorem: if the partial derivative $\partial f/\partial y$ exists and is bounded by $L$, then $|f(x,y_1)-f(x,y_2)|\le L|y_1-y_2|$ pointwise. The MVT is thus the bridge from differentiability of the vector field to the quantitative estimate the contraction needs.

• Fundamental theorems of calculus 02.04.04. Equivalence of the IVP $y^{\prime }=f(x,y)$ with the integral equation $y=Ty$ is the fundamental theorem of calculus, and it is what turns a differential problem into a fixed-point problem. The Picard operator, Gronwall's integrating factor, and the variation-of-constants formula 02.12.13 are all downstream of this single equivalence.

• Phase space, vector field, integral curve 02.12.01. Existence and uniqueness through every point are exactly what promote a vector field into a flow ${\varphi }_t$, a one-parameter family of maps moving states along integral curves. Without uniqueness there is no well-defined flow map; the Picard-Lindelof theorem is the existence certificate the geometric theory of ODEs stands on.

• Lyapunov stability, direct method 02.12.08. The continuous-dependence estimate $\Vert y(x;y_0)-y(x;z_0)\Vert \le \Vert y_0-z_0\Vert e^{Lx}$ is the prototype stability bound: nearby starts stay nearby, with rate controlled by the Lipschitz constant. Gronwall's inequality, proved here in full, is the engine behind virtually every Lyapunov-function estimate.

Historical & philosophical context Master

Cauchy, in his 1820s lectures on the calculus at the Ecole Polytechnique, gave the first existence theorems for $y^{\prime }=f(x,y)$ under the hypothesis that $f$ be analytic, constructing the solution as a power series. This was the seed, but analyticity is a heavy price. Picard 1890, in his Memoire sur la theorie des equations aux derivees partielles et la methode des approximations successives ^[Picard1890], replaced analyticity by continuity plus a Lipschitz condition and gave the iteration scheme bearing his name, proving convergence of the successive approximations to the unique solution. Lindelof 1894, in the Acta Societatis Scientiarum Fennicae ^{[Lindelof1894]}, refined the hypotheses and showed that the Lipschitz condition is the sharp dividing line: it is what one needs for uniqueness, over and above existence.

Peano 1890, in Demonstration de l'integrabilite des equations differentielles ordinaires ^[Peano1890], proved that continuity alone already guarantees existence, by a compactness argument (the lineage of the modern Euler-polygon and Arzela-Ascoli proof). The equation $y^{\prime }=\sqrt{|y|}$, $y(0)=0$, exhibited in the same paper, showed that Peano existence without Lipschitz genuinely permits non-uniqueness — a philosophical puncture in Laplacian determinism: a perfectly continuous law of evolution can fail to determine a unique future. Gronwall 1919, in the Annals of Mathematics ^{[Gronwall1919]}, introduced the inequality that now bears his name while studying dependence on a parameter, supplying the estimate that welded uniqueness, continuous dependence, and stability into a single quantitative framework.

Bibliography Master

@book{CoddingtonLevinson1955,
  author = {Coddington, Earl A. and Levinson, Norman},
  title  = {Theory of Ordinary Differential Equations},
  publisher = {McGraw-Hill},
  year   = {1955},
}

@book{Hartman1982,
  author = {Hartman, Philip},
  title  = {Ordinary Differential Equations},
  edition = {2nd},
  publisher = {Birkhauser},
  year   = {1982},
}

@book{Arnold1992,
  author = {Arnold, V. I.},
  title  = {Ordinary Differential Equations},
  publisher = {Springer-Verlag},
  year   = {1992},
}

@article{Picard1890,
  author = {Picard, Emile},
  title  = {Memoire sur la theorie des equations aux derivees partielles et la methode des approximations successives},
  journal = {Journal de Mathematiques Pures et Appliquees},
  volume = {4},
  year   = {1890},
  pages  = {145--210},
}

@article{Lindelof1894,
  author = {Lindelof, Ernst},
  title  = {Sur l'application de la methode des approximations successives aux equations differentielles ordinaires du premier ordre},
  journal = {Acta Societatis Scientiarum Fennicae},
  volume = {21},
  year   = {1894},
}

@article{Peano1890,
  author = {Peano, Giuseppe},
  title  = {Demonstration de l'integrabilite des equations differentielles ordinaires},
  journal = {Mathematische Annalen},
  volume = {37},
  year   = {1890},
  pages  = {182--228},
}

@article{Gronwall1919,
  author = {Gronwall, Thomas H.},
  title  = {Note on the derivatives with respect to a parameter of the solutions of a system of differential equations},
  journal = {Annals of Mathematics},
  volume = {20},
  number = {4},
  year   = {1919},
  pages  = {292--296},
}