The elementary contraction proof of an ODE often requires a short interval: its integral operator has a Lipschitz bound , which may not be below one on a longer interval. That restriction belongs to the chosen estimate, not necessarily to the equation. Ordered repeated integration introduces a factorial, allowing Picard iterates to converge on any fixed finite interval when the vector field is globally Lipschitz in the state.
Let with , and let be continuous. Assume one constant satisfies
The norm on is fixed; continuous curves have the corresponding supremum norm. The integral from to is oriented, so the same formulas cover times on either side of .
The Picard map and its powers
Definition 3.3.1
(Picard map)
For initial data , define on the complete space
Continuity of the integrand makes a self-map. A fixed point is a continuously differentiable solution of , , by the fundamental theorem of calculus; conversely every such solution is a fixed point.
The immediate estimate is . It may not be a contraction. Repeated integration nevertheless gives a sharper estimate for the powers of .
Lemma 3.3.2
(Factorial estimate for powers of the Picard map)
For all , all and all ,
In particular , so is a contraction as soon as .
Proof
(Induction over nested integrals)
For , the Lipschitz hypothesis gives . For the induction step integrates the previous estimate from to : the integral of is . For , reverse the limits and integrate from to . The factorial comes from these nested intervals. Since , for sufficiently large we have , so an iterate is a contraction even when is not.
Existence by successive approximations
Put and . The superscript is an iteration index, not a derivative. Let , which is finite by continuity.
Theorem 3.3.3
(Existence on the whole interval, with a tail bound)
Under the global Lipschitz hypothesis, the Picard iterates converge uniformly on to a solution of , . For every ,
where the factor is read as , including when . No assumption is made.
Proof
(Proof by a factorial majorant)
For ,
For the integral has norm at most . The same integration induction as in the lemma proves the other bounds. When , all differences from onward vanish, so only the first difference can be nonzero. Otherwise the supremum majorants form a convergent exponential series. Hence the telescoping sum of differences converges uniformly in the complete function space. Its limit satisfies : the original estimate proves continuity of even when . Summing the difference bounds from index onward, with , gives the tail bound. This controls exact Picard iterates. Rounding or numerical quadrature would add further errors not included in the bound.
Proof X-ray
A contraction can appear after iteration
The one-step estimate may be larger than one on a long interval. Repeated integration does not simply multiply that estimate: the ordered integration region contributes . The factorial majorant proves uniform convergence directly and also shows that a sufficiently high power of the Picard map is a contraction. Neither conclusion requires pretending that the one-step map is a contraction.
The edge case shows what the tail bound measures when there is no state dependence at all.
Worked example 3.3.4
(A state-independent field: one iterate is exact)
Take , and on with . Every works; take . Then , and every later iterate equals , because ignores its argument. So is reached after one step.
The tail bound agrees. Here . For only the term survives, giving ; the actual error is . For every term contains a positive power of , so the bound is , which is exact. Reading as is what keeps the case correct.
Uniqueness and a global consequence
Theorem 3.3.5
(Uniqueness and global existence)
On there is at most one solution with ; with the preceding theorem, the initial-value problem has exactly one solution on the whole interval . If is continuous on with one global state-Lipschitz constant , the initial-value problem has a unique solution defined for all real times.
Proof
(Uniqueness from the power estimate, then patching)
If are two fixed points, the power estimate gives
The factor tends to zero, forcing . Thus existence and uniqueness hold on the entire chosen finite interval, with no assumption . The same argument applies on any smaller closed interval containing , so solutions constructed on nested intervals agree on their overlaps.
For the global statement, apply the construction on for each . Solutions on nested intervals agree by uniqueness, so they patch to a unique solution for all real times. The solution need not be bounded as ; the conclusion is existence at every finite time.
Concept lensAnalytic
A weighted norm makes one step a contraction
Instead of iterating, one can change the norm. For put , which is equivalent to the supremum norm, so stays complete. Since , we get . The Picard map itself is then a contraction on the whole interval, and the exponential weight plays the role of the factorials.
Worked example 3.3.6
(Picard iterates of on )
For , , induction gives : integrating from gives . On , and the one-step bound is , so it does not establish contraction. The power factor equals for and first drops below one at , where it is ; thus is a contraction there.
Here , and the tail bound is
At the actual error is exactly this sum, so the bound is attained. Numerically it is about for , for and for ; nine iterates are therefore needed for a uniform error below . Enlarging the interval changes the bound but does not obstruct existence.
The explorer below compares the one-step factor with the power factor and the explicit tail bound with the actual supremum error of the Picard iterate, for on and three other presets.
Stability and residuals
Proposition 3.3.7
(Stability and residual bounds)
If solve the same equation on with initial values , then
Let be the solution with , and let have integral residual
Then .
Proof
(Repeated substitution of the integral inequality)
Repeatedly substitute the integral inequality for the difference of and . If , after steps one obtains
Letting yields . The inequality is valid for both time directions by the same oriented-integral argument.
For the residual bound, subtract the integral equation of from the definition of : . Hence satisfies , and the same repeated substitution gives . A small residual thus certifies a small error with an explicit amplification factor.
Worked example 3.3.8
(Two solutions of )
The field has global Lipschitz constant , since . Every solution therefore exists for all real times. Take and . The first solution is . For the second, , so
The proposition predicts . At the exact value is , against the bound ; at it is against ; at it is , well inside . The bound is nearly sharp for small , where , and it grows without limit although stays below . The estimate is valid at every time, but it is an upper bound, not a prediction of the actual separation.
Common mistake
(A small residual at one time is not enough)
The residual bound uses over the whole interval. A small residual at one time alone does not suffice. For , on , the curve has , which vanishes at ; yet , matching .
Why the global hypothesis matters
Counterexample mode
A locally Lipschitz equation can escape in finite time
For , , the maximal solution through the initial data is on . Its value tends to infinity as , so it cannot extend continuously through time 1. The field is locally Lipschitz but has no global Lipschitz constant on . Local existence and uniqueness do not imply global existence.
The global Lipschitz constant gives existence at every finite time; it says nothing about the size of the solution over an infinite time axis.
Common mistake
(Global existence does not mean boundedness)
A global solution need not converge to a finite limit as time grows. For , the solution exists for all finite and is unbounded as . Existence on every finite interval is different from a uniform bound over an infinite time axis.
Practice with solutions
Quick check
(For , , compute and give the general iterate.)
Solution · Answer
The iterates are , and ; in general .
Quick check
(Suppose , . Is the one-step estimate enough to prove that is a contraction? Why can its iterates still converge?)
Solution · Answer
No: the bound is , not less than one. The factorial bound on repeated integrals gives uniform convergence on the finite interval anyway; for example from onward, so is a contraction.
Quick check
(Two solutions have initial distance at most , with . Bound their distance at times with .)
Solution · Answer
The bound is .
Quick check
(On , an approximate curve has integral residual norm at most and the vector field has . Give an error bound.)
Solution · Answer
The residual estimate gives . It applies to the entire interval, not just at its endpoints.