The contraction proof of an ordinary differential equation does two jobs at once: it constructs a solution and proves that two solutions with the same initial value must agree. A continuous vector field need not provide the Lipschitz estimate used in that argument. Existence can nevertheless survive. Instead of showing that successive approximations are Cauchy, we can keep them in a compact family and extract a convergent subsequence.
This chapter develops that alternative for finite-dimensional systems. Euler polygons supply approximate trajectories; continuity makes their integral residuals small; Ascoli's theorem supplies a limit. The resulting solution need not be unique, and the construction need not recover every possible solution. We will keep these conclusions separate throughout.
Motivation
An Euler step freezes the velocity at the beginning of a short time interval. If the vector field is continuous, a small change in time and state changes that velocity only a little on a compact region. This suggests that the polygon approximately satisfies the integral equation.
Small residual alone must still be connected to convergence. On a compact time interval, a bounded vector field limits the amount a genuine integral trajectory can move. Approximate integral trajectories inherit this control up to their small residuals. That provides equicontinuity, the hypothesis which allows bounded families of curves to have uniformly convergent subsequences.
Uniqueness has a different role. If every convergent subsequence must approach the same solution, compactness upgrades subsequential convergence to convergence of the whole approximation sequence. A Lipschitz condition is one way to prove uniqueness; it is not the only way.
Definitions
Fix a positive integer , a time , an initial state and numbers . Use the Euclidean norm and the compact cylinder
Let be continuous. Choose with on . If , choose
if , choose any . Write .
Definition 3.4.1
(Classical solution and integral residual)
A solution on is a curve with values in such that and . Derivatives at the two endpoints are one-sided.
For any continuous curve , define its integral residual by
The residual is defined without requiring to be differentiable. If it is zero, continuity of the integrand and the fundamental theorem of calculus make a classical solution with initial value .
Definition 3.4.2
(Euler polygons)
For , set and . Starting at , recursively set
On , define
The induction below proves that all the evaluations lie in , so the recursion is legitimate. The polygon is continuous and piecewise linear; its derivative may jump at the nodes.
Theorem/proposition
Theorem 3.4.3
(Peano existence on a controlled interval)
Under the stated continuity and finite-dimensional hypotheses, there is at least one solution on , with
The Euler polygons have a uniformly convergent subsequence, and every uniformly convergent subsequence has a solution as its limit. No state-Lipschitz hypothesis or uniqueness assertion is included.
Theorem 3.4.4
(Compactness of consistent integral approximations)
Let be continuous, with and . Then every subsequence has a further uniformly convergent subsequence, and all such limits are solutions with initial state .
No common Lipschitz bound on the approximating curves is required: the integral equation up to a vanishing residual supplies the equicontinuity needed for compactness instead.
Corollary 3.4.5
(The solution family is compact; uniqueness gives full convergence)
The set of all solutions on taking values in and starting at is nonempty and compact in with the uniform norm. If has exactly one member , every consistent approximation sequence in the preceding theorem converges uniformly to . In particular, the whole Euler sequence converges whenever this solution is unique within the stated cylinder.
The family is also connected by the separate Kneser–Hukuhara theorem; a connectedness theorem of Mitschele, Theorem 2, treats a broader setting. Connectedness is not proved here. Compactness alone does not select a canonical solution or give a convergence rate to a chosen solution.
Proof sketch or proof idea
A small integral residual gives equicontinuity
Proof
(Proof of the compactness principle)
Write . For , subtraction of the two residual identities gives
This is an asymptotic modulus of continuity, not a Lipschitz estimate for each : the individual curves need not be Lipschitz. Fix . For all sufficiently large , the last term is less than . Choose a positive time-distance so that ; if , any positive works. This controls all the tail curves. The finitely many earlier continuous curves are uniformly continuous on , so shrinking if necessary controls them as well. The whole sequence is equicontinuous.
Its values lie in a bounded subset of , whose closure is compact. Apply the scalar Ascoli theorem to each coordinate and successively extract subsequences. There are only finitely many coordinates, so the final subsequence converges uniformly as a vector-valued curve. This explains the finite-dimensional step: a bounded set in a general infinite-dimensional normed target need not be relatively compact.
Let uniformly. The state ball is closed, so stays inside it, and the initial values give . Uniform continuity of on gives
Consequently the integrals converge uniformly, since their difference is bounded by times this supremum. Pass to the limit in the residual identity:
The integrand is continuous, so is and solves the equation. The same argument begins with any subsequence, proving the full compactness principle.
Proof X-ray
Consistency, compactness and identification are separate steps
A vanishing residual controls the approximation's error in the integral equation. The bounded field then gives equicontinuity, and Ascoli selects a uniform limit. Only after selection do uniform continuity and the integral identity identify that limit as a solution. None of these steps compares two different solutions.
Euler polygons remain in the cylinder
Proof
(Euler polygons and their frozen integral identity)
Assume that the nodes through have been constructed and satisfy . The velocity at that node is defined and has norm at most . For ,
Thus the next node also lies in the state ball, completing the induction. If , every polygon is simply constant. The bound on the velocities also gives across any number of cells.
For almost every , let be the left endpoint of its cell. The value assigned at the finitely many nodes does not affect the integral. The polygon satisfies exactly
This is not yet the desired integral equation: the time and the state on the right have both been frozen.
Uniform continuity makes the freezing error vanish
Proof
(Proof of Peano existence)
Define a joint modulus on the compact cylinder:
It is finite and tends to zero as . Within a cell,
Subtracting the desired integral from the exact frozen one yields
Apply the compactness principle. At least one subsequence converges to a solution, proving Peano existence. The polygon displacement estimate passes to the limit and gives the stated bound on .
This proof supplies a forward interval from the closed cylinder. If is continuous on an open time-state neighborhood of , choose a smaller compact cylinder on each side. Applying the forward theorem to the reversed field constructs the backward branch. The branches have the same value and derivative at , so they join to a solution on a two-sided neighborhood.
Compact solution families and a unique possible limit
Proof
(Proof of the corollary on solution families)
Every solution in is -Lipschitz and bounded in the fixed ball. Ascoli makes the family relatively compact. A uniform limit of its members satisfies the integral equation by uniform continuity of , so it is again in . Thus is closed and relatively compact, hence compact; Peano existence makes it nonempty.
If but a consistent sequence did not converge to , some subsequence would stay a fixed positive uniform distance from . The compactness principle would give a further subsequence converging to a member of , necessarily , a contradiction. This proves convergence of the whole sequence. It gives no quantitative rate by itself.
Concept lensAnalytic
Existence uses compactness; uniqueness identifies the limit
Without uniqueness, different approximate families may approach different exact solutions, and a fixed family need not approximate every solution. With uniqueness, all cluster points coincide, so precompactness forces convergence. A contraction estimate is not needed for this last implication.
Worked examples
Worked example 3.4.6
(A concrete interval and a residual bound)
Take , , on . The field is continuous and bounded by , so is allowed. No bounded state-Lipschitz constant exists near zero.
For real , . This follows from the triangle inequality and the subadditivity for nonnegative . Along an Euler cell, the time error is at most and the state error at most . Therefore
This proves consistency at an explicit rate and guarantees a convergent subsequence. It is not an error bound to a specified solution. Residual size and trajectory error are different quantities.
Worked example 3.4.7
(A family of waiting-time solutions and an Euler choice)
For , , consider . For every ,
is a solution. The derivative is zero at the joining time, so the equation holds there too. These curves stay in , where supplies the common interval .
They are all the solutions on this interval. The field is nonnegative, so a solution from zero is nonnegative and nondecreasing. Its zero set is an initial interval ending at some . If it ever becomes positive, then on the positive interval . Integrating and using continuity at gives . The all-zero solution corresponds to .
Every unperturbed left-endpoint Euler polygon from zero is identically zero: a zero node always produces another zero node. Thus the Euler sequence converges, but only to one member of the solution family. Convergence of an approximation does not prove uniqueness and does not promise access to every solution.
Worked example 3.4.8
(A non-Lipschitz field can still have uniqueness)
Let , with . It is continuous and nonincreasing but is not Lipschitz near zero. For two solutions ,
Equal initial values force equality for all later times in their common interval. Forward uniqueness holds despite the failed Lipschitz condition. For initial value , the forward solution is
It reaches zero and remains there, with a continuous derivative at the hitting time. On any interval supplied by the Peano construction, uniqueness upgrades Euler subsequence convergence to convergence of the whole sequence. Absence of a Lipschitz estimate leaves uniqueness undecided; it does not imply its failure.
Counterexample mode
A bounded discontinuous field can defeat the limit argument
On , set and start at zero. All nodes of the uniform Euler grid are rational, so every Euler polygon is . They are bounded and equicontinuous and converge uniformly.
Their limit is not a classical solution: its derivative is one at irrational times too, whereas the field is zero there. No solution could have derivative equal to this field on an interval, because a continuous derivative cannot be one on the rationals and zero on the irrationals. For this discontinuous example, read the residual formula using the Lebesgue integral. That integral of is zero, so ; consistency fails completely. Boundedness and compactness alone do not replace continuity of the field. An almost-everywhere differential equation is a different formulation.
Common mistakes
Common mistake
(A polygon is not a classical solution)
A polygon is generally not a classical solution: its slopes jump at mesh nodes and its velocity is frozen. The proof uses its exact frozen integral identity and then estimates the residual. Differentiating a uniform limit of polygons would not be a valid substitute.
Common mistake
(Ascoli needs equicontinuity, not only bounds)
Ascoli requires equicontinuity as well as bounded values. Here bounded velocity gives it directly for Euler polygons, and vanishing integral residuals give it for more general approximations. The finite-dimensional compactness of the state ball is also a real input.
The guaranteed time interval is a local estimate, not a maximal lifespan. It depends on a chosen cylinder and its velocity bound. Compactness of the solution family in that cylinder does not establish a unique maximal solution, a single-valued flow, or differentiable parameter dependence.
Common mistake
(A small residual is not a small trajectory error)
A small residual need not mean a small distance to one preselected solution. In the waiting-time example every has residual zero, although different members can be far apart. A uniqueness or stability argument is needed for a conclusion about one selected trajectory.
Summary
Continuous finite-dimensional fields admit local solutions because controlled Euler polygons have vanishing integral residuals and uniformly convergent subsequences. Continuity identifies the limits through the integral equation. The full family of solutions on the controlled interval is compact. Uniqueness, when proved separately, then forces every consistent approximation to have the same limit.
Exercises
Quick check
(Suppose on that , where and . Prove an explicit Euler residual bound using . Explain why it is not automatically a trajectory-error bound.)
Solution · Guided solution
Within an Euler cell, the two arguments differ by at most in time and in state. Integrating the stated estimate gives . No comparison with a chosen exact solution has been made. Without uniqueness, another exact solution can have residual zero and still differ from that chosen solution. Even with uniqueness, this compactness proof alone supplies convergence rather than a quantitative stability estimate.
Quick check
(For , let . Show that these fields converge uniformly to , even on the whole real line. Verify that solves the perturbed initial-value problem from zero for . Compare its limit with the unperturbed Euler limit.)
Solution · Guided solution
For , , so . Since and for , , with . On every fixed , it converges uniformly to , with error . The exact-field Euler polygons converge to zero instead. Both limits solve the same unperturbed initial-value problem. Uniform convergence of fields does not select a unique branch when the limiting equation is nonunique.
Quick check
(Let uniformly on the same compact cylinder, with all fields continuous. Suppose classical solutions of are already given, and . Prove subsequential uniform convergence to a solution of the limiting equation. State the extra condition giving full convergence. Do not assume common existence merely from the notation.)
Solution · Guided solution
Write the equation for in integral form. Its residual relative to is , so . The curves stay in the same state ball and their initial values converge, exactly as required by the compactness principle. Every subsequence has a further uniform subsequence converging to a solution of with initial state . If that solution is unique among curves remaining in the cylinder, the entire sequence converges. The hypotheses explicitly provide the curves on the common interval; existence there has not been inferred from uniform convergence alone.
Quick check
(For the waiting-time equation on , compute the set of possible values at time . Prove directly that its full solution family is compact by estimating in terms of .)
Solution · Guided solution
The classification gives . As ranges over , these values fill : for a value in that interval choose , and zero is also obtained by every . For , the positive-part function is 1-Lipschitz and its values here are at most one, so . Thus is continuous into the uniform-norm space. Its image of the compact interval is compact and, by the classification, is the entire solution family. This proves compactness in this example without using a general connectedness theorem.
Quick check
(On , set for integers . Show that this bounded family has no uniformly convergent subsequence. If a continuous field satisfies on , prove that the residuals of these curves cannot tend uniformly to zero.)
Solution · Guided solution
Pointwise, tends to zero except at , where it equals one. Every subsequence has that same discontinuous pointwise limit, so none can converge uniformly to a continuous function. For the residual associated with the given field, put and . The two curve values are one and zero. Subtracting the residual identities gives . Hence , whose lower limit is at least one half. These spikes cannot be consistent integral approximations to any fixed continuous bounded field on the stated cylinder.