Existence and Uniqueness Theorems for ODEs

These theorems are the silent clause under every modern computation that trusts a differential equation to have an answer. Every adaptive ODE solver — the Dormand–Prince engine inside SciPy and MATLAB's `ode45`, and the stiff `ode15s`/BDF codes for chemical-kinetics and circuit models — rests on a Lipschitz constant: it bounds the truncation error through a discrete Gronwall inequality and certifies that the exact solution being approximated exists at all. In weather prediction, the Lorenz system's sharp Gronwall separation is the rigorous face of the butterfly effect — solutions are unique yet diverge exponentially, so forecasts are deterministic but bounded in horizon. The machine-learning frontier has made the theorem newly fashionable: Neural ODEs (Chen and collaborators, 2018) parameterize $f(x,y)$ by a neural network and lean on Lipschitz control to keep the forward integration and its adjoint pass well-posed. And the deepest open question is this same theorem promoted to PDEs: whether the three-dimensional incompressible Navier–Stokes equations admit unique, globally smooth solutions is a Clay Millennium Prize problem — the billion-dollar sibling of the scalar dichotomy proved here.

Read $f(x,y)$ as a slope field filling the plane: at every point it posts an arrow telling a passing trajectory which way to head. Existence asks whether you can thread a smooth curve through this thicket starting at $(x_0,y_0)$; uniqueness asks whether two such curves can ever cross or graze. The Lipschitz condition caps how quickly the posted slope changes vertically, and that cap is exactly what forbids two integral curves from merging. The first figure makes it concrete: for $y'=y$ the slopes fan out gently, and a single exponential is threaded cleanly through the initial point.

Picard iteration is that slope field made dynamical. Start with the flat guess $y_0(x)=y_0$, then repeatedly feed the current guess into the field and integrate to get a better one. Each pass nudges the curve to honour the slopes it passes, and the Lipschitz contraction shrinks the error by a fixed factor every time — geometric convergence. The second figure shows this for $y'=y$, $y(0)=1$: the iterates are literally the partial sums of $e^x$, marching upward toward the exponential.

The failure of uniqueness in $y'=3y^{2/3}$ is geometrically vivid. Along $y=0$ the posted slope is zero, so a trajectory can ride that line for an arbitrary stretch and then peel off upward as a cubic. The slope $\partial f/\partial y=2y^{-1/3}$ blows up at $y=0$, so the field gives no instruction sharp enough to pin down when to leave — hence a continuum of solutions. Uniqueness and continuous dependence are two faces of one coin: if nearby data must give nearby solutions (Gronwall control), then identical data give identical solutions — Hadamard well-posedness, the mathematical content of determinism.

The starting point is the initial-value problem: a first-order equation together with one data point the solution must pass through.

where: $f(x,y)$ — the prescribed slope function on a region $R$ of the $(x,y)$-plane; $(x_0,y_0)$ — the initial point the solution threads through; $y'=\mathrm{d}y/\mathrm{d}x$ — the slope demanded at each $x$. Reading: the equation posts a slope everywhere, the data pins the curve at one point. The first structural move is to integrate. The fundamental theorem of calculus folds (1) and its initial condition into a single integral equation, which asks only that $y$ be continuous to make sense.

where: $t$ — the dummy integration variable. Reading: a solution of (2) satisfies both the ODE (differentiate) and the data (set $x=x_0$), so (1) and (2) are equivalent — but (2) is a fixed-point equation, ready for iteration. Uniqueness lives or dies on how sharply $f$ varies in its second slot. The Lipschitz condition demands that the slope not change faster than linearly as you move vertically — a uniform, checkable cap.

where: $L$ — the Lipschitz constant, uniform over $x$; $y_1,y_2$ — any two ordinates. Reading: if $\partial f/\partial y$ exists and is bounded by $L$, (3) holds automatically by the mean value theorem — the everyday sufficient test. Equation (2) defines an operator that eats a trial function and returns a better one; a solution is precisely a fixed point of it, and iterating generates the Picard approximations.

where: $T$ — the Picard integral operator on continuous functions; $y_n$ — the $n$-th approximation, from the constant guess $y_0(x)=y_0$. Reading: solving the ODE becomes solving $y=Ty$, each application of $T$ threading the current guess back through the slope field. The theorem is local: it guarantees a solution only on an interval whose half-width is fixed by how big $f$ can get inside a chosen rectangle. Confining the iterate to the rectangle is what forces the half-width down.

where: $R$ — the closed rectangle on which $f$ is bounded; $M$ — the bound on $|f|$; $h$ — the guaranteed half-width of existence. Reading: over $|x-x_0|\le h$ the iterate cannot escape the rectangle, so $T$ maps the working space into itself. Finally, the quantitative heart of uniqueness: Gronwall's inequality converts an integral self-bound on a nonnegative function into an explicit exponential ceiling — the one-line engine behind uniqueness, continuous dependence, and error estimates alike.

where: $u(x)\ge0$ — a continuous function (the gap between two solutions); $a\ge0$ — a constant offset; $b\ge0$ — a constant rate (the Lipschitz $L$). Reading: with $a=0$ the ceiling is zero, so a function that bounds itself by its own running integral must vanish — the trick that kills the difference of two solutions.

Everything about the IVP hinges on the regularity of $f$ in its second argument — the table reads across the spectrum from a global Lipschitz bound down to bare continuity.

1 · Sizing the interval of existence (fully worked)

Consider $y'=x^2+y^2$ with $y(0)=0$. On $R=\{|x|\le1,\ |y|\le1\}$, find the guaranteed interval of existence from Picard–Lindelöf, and the Lipschitz constant that certifies uniqueness.

Step 1. Bound $f$ on the rectangle. Since $f=x^2+y^2$ grows with both $|x|$ and $|y|$, its maximum on $R$ sits at a corner: $M=\max_R|x^2+y^2|=1^2+1^2=\mathbf{2}$ — the fastest the solution can climb inside $R$.

Step 2. Apply (5) with $a=1$, $b=1$: $h=\min(a,\,b/M)=\min(1,\,1/2)=\mathbf{1/2}$, so a unique solution is guaranteed at least on $|x|\le\tfrac12$. The binding constraint is $b/M$, not $a$ — the solution could reach the top of the rectangle before $x$ reaches $1$, so the interval is cut back to protect against escape.

Step 3. Certify uniqueness. On $R$, $\partial f/\partial y=2y$, so $L=\max_R|2y|=2$; being continuous and bounded, $f$ is Lipschitz in $y$ and the hypotheses are met in full.

Step 4. Read the caveat. The naive contraction factor is $q=Lh=2\cdot\tfrac12=1$, exactly at the boundary — the plain argument would need $h<\tfrac12$, but the sharpened weighted-norm theorem secures existence on $|x|\le\tfrac12$ anyway. Guaranteed interval $|x|\le 1/2$, Lipschitz constant $L=2$, bound set by $b/M$ not $a$. (This Riccati equation has no elementary closed form — exactly why an a priori existence interval is valuable.)

2 · Picard iteration reconstructs the exponential (fully worked)

Solve $y'=y$, $y(0)=1$ by Picard iteration and check the convergence at $x=1$.

Step 1. Set up operator (4): $y_{n+1}(x)=1+\int_0^{x} y_n(t)\,\mathrm{d}t$, from $y_0(x)=1$.

Step 2. Iterate: $y_1=1+\int_0^x 1\,\mathrm{d}t=1+x$; $y_2=1+\int_0^x(1+t)\,\mathrm{d}t=1+x+\tfrac{x^2}{2}$; $y_3=1+x+\tfrac{x^2}{2}+\tfrac{x^3}{6}$. The pattern is unmistakable: $y_n(x)=\sum_{k=0}^{n}\dfrac{x^k}{k!}$, the partial sums of $e^x$.

Step 3. Take the limit: $y_n\to\sum_{k=0}^{\infty}x^k/k!=e^x$, the exact solution — existence constructively, uniqueness because $f=y$ is globally Lipschitz ($L=1$).

Step 4. Check numerically at $x=1$: the iterates give $y_1=2.000$, $y_2=2.500$, $y_3\approx2.667$, $y_4\approx2.708$, $y_5\approx2.717$, closing on $e=2.71828$. The errors shrink even faster than the guaranteed geometric $q^n/(1-q)$ ceiling (a factorial tail). The Picard sequence is the Taylor series of $e^x$, assembled one integration at a time.

3 · When uniqueness fails (qualitative)

Show that $y'=3y^{2/3}$, $y(0)=0$ has more than one solution and locate the culprit.

Step 1. Verify two solutions by substitution. The flat function $y\equiv0$ gives $y'=0=3\cdot0^{2/3}$. The cubic $y=x^3$ gives $y'=3x^2$, while $3y^{2/3}=3(x^3)^{2/3}=3x^2$ — a match. Both satisfy the same data $y(0)=0$.

Step 2. Build the whole family. For any $c\ge0$, the delayed cubic $y=(x-c)^3$ for $x\ge c$, glued to $y\equiv0$ for $x\le c$, is continuously differentiable and solves the IVP — infinitely many solutions, one per launch time $c$.

Step 3. Name the culprit. Uniqueness fails because $f=3y^{2/3}$ is not Lipschitz at $y=0$: $\partial f/\partial y=2y^{-1/3}\to\infty$ as $y\to0$. Where the slope-in-$y$ diverges the solution can loiter on the axis and then depart, so the field cannot decide when. Peano still guarantees existence; only uniqueness is lost.

The existence–uniqueness theory was built by several hands — Cauchy laid the foundation, Lipschitz set the load-bearing condition, Peano probed how little you can get away with — but the working method physicists run carries one name above the others.

Texts. Stone and Goldbart, Mathematics for Physics, the chapter on ODEs and existence theory · Arfken, Weber and Harris, Mathematical Methods for Physicists, 7th ed., Ch. 7 · Riley, Hobson and Bence, Mathematical Methods for Physics and Engineering, 3rd ed., Ch. 14 · Coddington and Levinson, Theory of Ordinary Differential Equations, Ch. 1–2 · Hartman, Ordinary Differential Equations, Ch. II–III. Key numerical anchors: the contraction estimate gives $\|Ty-Tz\|_\infty\le Lh\|y-z\|_\infty$ with $q=Lh<1$; $y'=3y^{2/3}$ is non-unique via $y\equiv0$, $y=x^3$, and $y=(x-c)^3$; $y=1/(1-x)$ solves $y'=y^2$, $y(0)=1$ and blows up at $x=1$; on $|x|,|y|\le1$ the example gives $M=2$, $h=\tfrac12$, $L=2$; the Picard iterates for $y'=y$, $y(0)=1$ are the partial sums of $e^x$, with $y_3(1)=2.6\overline{6}$ against $e=2.71828$; and $a=0\Rightarrow u\equiv0$ in Gronwall.

Builds on. First-Order Ordinary Differential Equations · Integrating Factors and Exact Equations. Leads to. Linear Second-Order ODEs · The Wronskian and Linear Independence · Power Series Solutions · The Method of Frobenius.

Ask the tutor: Ask me to run Picard iteration on your own $y'=f(x,y)$, size its interval of existence with $h=\min(a,b/M)$, or diagnose whether a given equation branches — and I will walk each step with you.