Existence and Uniqueness (Picard-Lindelof)

The Picard iteration scheme is not merely a proof technique; it is a genuine computational method, closely related to the numerical integration schemes used to simulate ODEs on a computer, and it directly generalizes to integral and integro-differential equations arising in physics and engineering. In control theory, well-posedness (existence, uniqueness, continuous dependence on initial data) established via Picard–Lindelöf is the starting point for proving that feedback-controlled systems behave predictably, and the Lipschitz constant $L$ appears explicitly in stability margins and in the exponential bound of the continuous-dependence corollary above. In numerical analysis, the same contraction argument underlies convergence proofs for implicit numerical schemes (e.g. backward Euler) applied to stiff equations. In stochastic analysis, an analogous Lipschitz condition on drift and diffusion coefficients is precisely the hypothesis under which stochastic differential equations admit unique strong solutions, via a stochastic version of Picard iteration. In dynamical systems, uniqueness of solutions through every point is the structural fact that makes the flow map $\Phi_t$ well-defined and lets one speak coherently of trajectories, invariant sets, and phase portraits, ideas that recur when passing to Systems of First-Order Linear ODEs and to Second-Order Linear Equations (which reduce to first-order systems for exactly this reason). In partial differential equations, generalizations such as the Cauchy–Kovalevskaya theorem and fixed-point arguments in Banach spaces for nonlinear PDEs (e.g. via the contraction mapping principle in Sobolev spaces) are direct philosophical descendants of the same iterative strategy.

Ask the tutor: Can you show me step by step why the interval length $h=\min(a,b/M)$ can be forced smaller than $a$, using a specific numeric example?

Think of the IVP not as a differential statement but as an integral equation. If $y$ solves $y' = f(t,y)$, $y(t_0)=y_0$, then integrating both sides from $t_0$ to $t$ gives

Conversely, any continuous $y$ satisfying this integral equation is automatically differentiable (by the Fundamental Theorem of Calculus, since $f(s,y(s))$ is continuous) and solves the IVP. This equivalence is the entire strategy: turn a differential problem into a fixed-point problem for the map

Finding a solution to the IVP is the same as finding a function $y$ with $Ty = y$. Picard's idea is to build such a fixed point by brute-force iteration, starting from the crudest possible guess $y_0(t) \equiv y_0$ (the constant function), and repeatedly applying $T$:

Each $y_n$ is an explicit, computable function — a polynomial, if $f$ is polynomial. The Lipschitz condition is precisely what makes consecutive iterates get closer together geometrically fast, so that the sequence converges uniformly to a genuine fixed point, which is the solution. Geometrically, on the $(t,y)$-plane the direction field $f(t,y)$ assigns a slope to every point; a solution curve is a curve everywhere tangent to this field. Lipschitz continuity says the field cannot vary in the $y$-direction "too fast" — nearby heights get nearby slopes — which is exactly the condition preventing two different curves from touching and then splitting apart, since a splitting point would require the field to distinguish two coincident states.

Example 1 (Picard iterates for a genuinely asymmetric linear equation). Solve $y' = 2t + y$, $y(0) = 1$, using Picard iteration, and verify against the exact solution. Here $f(t,y) = 2t + y$ is continuous everywhere and satisfies $|f(t,y_1)-f(t,y_2)| = |y_1-y_2|$, so $L=1$ works on all of $\mathbb{R}^2$: the hypotheses hold on any box, giving a solution on some interval around $t=0$ (in fact, because $f$ is linear in $y$ with continuous coefficients, the solution extends to all of $\mathbb{R}$, though the local theorem alone only promises a short interval). Starting iterate: $y_0(t) = 1$.

Compare with the exact solution. Solving the linear equation $y' - y = 2t$ by an integrating factor $e^{-t}$:

Imposing $y(0)=1$: $1 = -2 + C$, so $C = 3$, giving

Its Taylor series at $t=0$ is $3\left(1 + t + \tfrac{t^2}{2} + \tfrac{t^3}{6} + \tfrac{t^4}{24}+\cdots\right) - 2t - 2 = 1 + t + \tfrac32 t^2 + \tfrac12 t^3 + \tfrac18 t^4 + \cdots$. The Picard iterate $y_3(t) = 1 + t + \tfrac32 t^2 + \tfrac12 t^3 + \tfrac{1}{12}t^4$ agrees with the exact series through the $t^3$ term and is already close in the $t^4$ term ($\tfrac1{12}$ vs. the true $\tfrac18$); each further iterate reproduces one more correct order of the Taylor series, exactly as the geometric error bound $(\dagger)$ predicts.

Example 2 (Checking hypotheses carefully on an explicit box, and computing $h$). Consider $y' = y^2 - t$, $y(0) = 0$. Take the box $R = \{|t| \le 1,\ |y| \le 1\}$ (so $a=1$, $b=1$, centered at $(t_0,y_0)=(0,0)$). Continuity and the bound $M$: $f(t,y) = y^2 - t$ is a polynomial, hence continuous everywhere, in particular on $R$. On $R$, $|y^2| \le 1$ and $|t| \le 1$, so

(This is a valid, if not sharp, bound — sharper analysis would give $M=1$ at the corners $(\pm1,\pm1)$ where $y^2-t$ reaches $\pm 2$ only when $y^2=1,t=-1$ giving $2$, so $M=2$ is in fact attained and is sharp.) Lipschitz condition: for $(t,y_1), (t,y_2) \in R$,

using $|y_1|,|y_2|\le 1$ on $R$. So $L = 2$ works. (This also follows from $\partial f/\partial y = 2y$, bounded by $2$ in absolute value on $R$, via the mean value theorem — consistent with the direct computation.) Computing $h$: $h = \min\{a, b/M\} = \min\{1, 1/2\} = \tfrac12$.

Conclusion: the theorem guarantees a unique solution on $[-\tfrac12, \tfrac12]$ — a genuinely local guarantee, since nothing in this argument extends the interval to all of $[-1,1]$ even though $R$ itself was chosen with $a=1$. This illustrates the typical situation: the interval length $h$ produced by the theorem depends on the size of the box through $M$, and enlarging $b$ to get more room in $y$ can shrink $h$ because $M$ may grow faster than $b$. A cleverer choice of box (e.g., a thinner one in $y$) can sometimes improve $h$, but Picard–Lindelöf itself gives no claim about the true maximal interval of existence — that requires a separate continuation argument.