The Picard-Lindelöf Theorem: Existence and Uniqueness for ODEs

The precise statement

Let D ⊂ ℝ × ℝⁿ be an open set containing the point (t₀, y₀), and let f : D → ℝⁿ be a function. Suppose:

• Continuity: f is continuous on D;
• Lipschitz in y: there exists a constant L ≥ 0 such that |f(t, y₁) − f(t, y₂)| ≤ L·|y₁ − y₂| for all (t, y₁), (t, y₂) in D.

Then there exists h > 0 such that the initial value problem

y′(t) = f(t, y(t)),   y(t₀) = y₀

has a unique solution y(t) on the interval [t₀ − h, t₀ + h]. Concretely, if f is bounded by M and Lipschitz-L on a closed box R = {|t − t₀| ≤ a, |y − y₀| ≤ b}, one may take h = min(a, b/M). The result holds verbatim for systems (y ∈ ℝⁿ), hence for higher-order scalar ODEs after reduction to first-order form. Note the theorem is local: it promises a solution only on a small interval, not for all time.

The picture: a vector field you must follow

Think of f as prescribing, at every point (t, y) of the plane, a little arrow of slope f(t, y). A solution is a curve that is everywhere tangent to these arrows — it flows along the field. The initial condition nails down one point the curve must pass through.

Existence asks: can you always thread a curve through the arrows starting from (t₀, y₀)? Uniqueness asks: could two distinct curves start at the same point and later diverge? The Lipschitz condition controls how fast neighboring arrows can differ. If arrows change too abruptly with y, two solution curves can peel apart or merge — uniqueness fails. The Lipschitz bound |f(t,y₁) − f(t,y₂)| ≤ L·|y₁ − y₂| forbids exactly this by capping the rate at which the field varies vertically.

The number h = min(a, b/M) is the geometric guarantee that the curve, moving with slope at most M, cannot escape the box R before time h elapses — so our estimates remain valid throughout.

The key idea: turn the ODE into a fixed point

The engine of the proof is a change of viewpoint. Integrating y′ = f(t, y) from t₀ shows a differentiable function y solves the IVP if and only if it satisfies the integral equation

y(t) = y₀ + ∫ from t₀ to t of f(s, y(s)) ds.

Define the Picard operator (T y)(t) = y₀ + ∫ from t₀ to t of f(s, y(s)) ds acting on continuous functions. A solution is precisely a fixed point T y = y. Now equip the space of continuous functions on [t₀ − h, t₀ + h] with the sup-norm, making it a complete metric space (a Banach space). One computes that T is a contraction: ‖T y₁ − T y₂‖ ≤ (L·h)·‖y₁ − y₂‖, so choosing h small enough that L·h < 1 makes the contraction constant < 1. The Banach fixed-point theorem then delivers a unique fixed point. Equivalently, the Picard iterates y₀, y₁ = T y₀, y₂ = T y₁, … converge uniformly to the solution — this is the constructive heart of the argument.

Worked example: recovering the exponential

Take the simplest nontrivial IVP: y′ = y, y(0) = 1, so f(t, y) = y, which is globally Lipschitz with L = 1. The Picard iteration starts from the constant y₀(t) = 1 and applies T y = 1 + ∫₀ᵗ y(s) ds:

• y₁(t) = 1 + ∫₀ᵗ 1 ds = 1 + t;
• y₂(t) = 1 + ∫₀ᵗ (1 + s) ds = 1 + t + t²/2;
• y₃(t) = 1 + t + t²/2 + t³/6;
• in general yₙ(t) = ∑ from k=0 to n of tᵏ/k!.

These are exactly the partial sums of the Taylor series, so yₙ(t) → ∑ tᵏ/k! = eᵗ. The iteration builds the exponential from scratch, and the theorem certifies eᵗ is the one and only solution. This is the canonical illustration: Picard iteration reconstructs the power series of the answer term by term, and the contraction estimate is what guarantees the series converges.

Why the hypotheses matter — sharp counterexamples

Drop the Lipschitz condition and uniqueness collapses. The textbook counterexample is y′ = √|y| = |y|^(1/2), y(0) = 0. Here f(y) = |y|^(1/2) is continuous but not Lipschitz at y = 0 (its difference quotient blows up). Two solutions coexist: the trivial y(t) ≡ 0, and y(t) = t²/4 for t ≥ 0 — in fact infinitely many, since the curve may rest at 0 for any duration before lifting off. Peano's theorem still gives existence (continuity alone suffices), but Picard-Lindelöf's Lipschitz hypothesis is exactly what buys uniqueness.

Drop even continuity and existence can fail. And the theorem is only local: y′ = y², y(0) = 1 has the unique solution y = 1/(1 − t), which escapes to +∞ at t = 1 despite f being smooth — a finite-time blow-up. Global existence needs an extra bound (e.g. linear growth in y). The completeness of the function space is likewise essential: Banach's fixed-point theorem is false without it, and Osgood's condition shows the Lipschitz bound can be relaxed to ∫₀ dr/ω(r) = ∞ while retaining uniqueness.

Significance: the license to solve ODEs

Picard-Lindelöf is the foundational well-posedness result that underwrites all of differential equations, dynamical systems, and mathematical physics. Every time one writes "the solution" to Newton's equations, a chemical rate law, an electrical circuit, or a population model, one is implicitly invoking this theorem — without uniqueness, "the" trajectory of a deterministic system would be meaningless.

Its consequences ripple outward. It justifies the flow of a vector field (the map (t, y₀) ↦ y(t)), the starting point of dynamical systems theory and Lyapunov stability. It grounds numerical methods (Euler, Runge-Kutta) — you can only approximate a solution that provably exists. The same contraction-mapping machinery generalizes to Banach-space-valued ODEs, underpinning the Cauchy-Kovalevskaya theory for PDEs and semigroup methods. And the Lipschitz-vs-merely-continuous dichotomy it exposes — uniqueness as a smoothness phenomenon — recurs throughout analysis, from characteristics of PDEs to the theory of stochastic differential equations, where an analogous Lipschitz condition secures pathwise uniqueness.

Existence and uniqueness theorems for the IVP y′ = f(t, y), y(t₀) = y₀, ordered by hypothesis strength.

Theorem	Hypothesis on f	Conclusion	Uniqueness?
Peano	f continuous only	At least one local solution exists	No — can fail
Picard-Lindelöf	f continuous, Lipschitz in y	Unique local solution exists	Yes
Osgood	f continuous, |f(t,y₁)−f(t,y₂)| ≤ ω(|y₁−y₂|) with ∫₀ dr/ω(r) = ∞	Unique local solution	Yes (weaker than Lipschitz)
Carathéodory	f measurable in t, continuous & Lipschitz in y	Unique absolutely continuous solution	Yes (allows discontinuity in t)
Global / continuation	Lipschitz + f defined on all of ℝⁿ (or linear growth)	Solution extends to all of ℝ	Yes