Solving an equation is often only the beginning. If the prescribed data change, does its solution move continuously? Can its first-order change be computed without solving the whole problem again? For a differential equation, another question comes first: will all nearby problems still have solutions at the time where we want to compare them?
The local contraction theorem constructs a solution on a short interval. Here the emphasis shifts to the dependence of that solution on its data, and to the conditions that let a short solution continue. The implicit function theorem then reappears in a useful new role: it converts a smoothly varying trajectory into a smoothly varying time at which a constraint is met.
Motivation
Consider an equation , where is prescribed and is unknown. An explicit formula for may be unavailable, while a nearby solution and its derivative can still be determined. The decisive question is whether a small change in produces enough independent change in the constraint to compensate for a small change in .
For an initial-value problem, the unknown is an entire curve. Perturbing the initial state changes the curve at once; perturbing a parameter changes its velocity throughout the interval. A useful sensitivity formula must account for both effects. Merely differentiating an equation containing an unknown curve is not a proof that the curve is differentiable in the parameter. We first establish a common existence interval and an error estimate, then justify the differentiation by controlling its remainder. What we obtain is a linear approximation to the actual solution, not merely another equation that is consistent with formal manipulation.
Definitions
All vector spaces below are finite-dimensional, with Euclidean norms and induced matrix norms. Let be an open interval, and open sets, and consider
Unless stated otherwise, is continuous, its state and parameter derivatives exist, and both derivatives are jointly continuous in . No time derivative of is required. These assumptions imply local Lipschitz control in the state and parameter on sufficiently small closed convex rectangles.
Definition 3.2.1
(Solution map and sensitivity matrices)
Where the initial-value problem has its unique solution, write it as , keeping the initial time fixed. When the indicated derivatives exist, define
The first matrix describes sensitivity to initial state; the second describes sensitivity to a parameter in the vector field. Differentiability here concerns the actual solution map on a common time interval, not a separately chosen formal solution for each parameter.
Definition 3.2.2
(Maximal solution and continuation)
For fixed , an extension of a solution is a solution on a larger interval that agrees with it wherever the original solution is defined. A maximal solution admits no proper extension inside the time-state domain . Its interval is open and contains . A finite endpoint need not mean the norm of the state becomes infinite: the state can approach a boundary point excluded from .
Theorem/proposition
Theorem 3.2.3
(Implicit branches and their linear response)
Let be , with and invertible. In a product neighborhood, all solutions have the unique form , where is and
This is a local statement about the selected branch. It neither excludes distant branches nor guarantees a branch through a point where the state derivative becomes singular.
Lemma 3.2.4
(Gronwall's integral estimate)
If a nonnegative continuous function on satisfies
then throughout that interval.
Theorem 3.2.5
(Differentiable dependence and variational equations)
Under the stated assumptions on , every base initial condition has a neighborhood in and a common time interval about on which the solution map is in . Along a solution, let
Then its sensitivities are the unique solutions of
The same conclusion holds on any compact time interval inside a fixed reference solution's existence interval, after the data neighborhood is made sufficiently small.
Concept lensAnalytic
Parameter sensitivity as transported velocity changes
Because solves with , it is invertible at every time: its determinant satisfies , so . The inhomogeneous equation for then has the explicit solution
Differentiating in gives and , so this is the same matrix. Read this way, a parameter change adds the velocity at each time , and carries that infinitesimal change of state from time to time , exactly as it would carry a change of state made at time . For this gives , the value used in the second worked example below.
Theorem 3.2.6
(Continuation within the time-state domain)
For fixed , local uniqueness gives a unique maximal solution. If its finite right endpoint belongs to , and a tail of its graph is contained in a compact subset of , the solution extends past . Thus this compact-containment situation cannot occur for a genuinely maximal finite endpoint. The analogous statement holds at a finite left endpoint.
Corollary 3.2.7
(A transverse constraint determines a local event time)
Let be . Keep the initial state fixed. Suppose
at that event, with inside the reference solution's interval. There is a unique nearby root time , and
with the quantities evaluated along the event. This is a local root-time statement, not automatically a theorem about the first event after .
Proof sketch or proof idea
Implicit constraints and an integral bound
Proof
(Proofs of the implicit-branch formula and Gronwall's estimate)
For the implicit theorem, apply the inverse function theorem to . Its derivative is block triangular, with diagonal blocks and . The local inverse evaluated at therefore gives . Differentiating yields the formula. This uses the inverse theorem already proved by contraction; it does not assume an explicit solution of the constraint. If is , repeated use of the derivative formula gives a branch, since matrix inversion is smooth on the invertible matrices.
For Gronwall, set . Then and , so . Hence . This also works when ; dividing by would unnecessarily exclude the case used for uniqueness.
Establish a common interval before comparing curves
Proof
(A common interval and continuous dependence)
Choose a closed time-state-parameter rectangle around the base data, contained in . On it let , and . If its state ball has radius , restrict initial states to the concentric ball of radius and choose with and . Picard iteration then preserves the same state ball for every such initial state and every parameter in the chosen parameter ball. It gives a common interval .
For two of these solutions, subtract their integral equations. On forward time, the state difference is bounded by the initial difference, the integral of times the state difference, and times the parameter difference. Gronwall gives, for ,
For backward time use and the reversed differential equation. This estimate proves continuous dependence, uniformly over the common interval. Constants must be valid on the whole comparison rectangle; a derivative bound only along one trajectory would not yet control a neighboring trajectory.
Justify the variational equations
Proof
(Proof of differentiable dependence)
Fix base data and write . The preceding estimate gives . Uniform continuity of the state and parameter derivatives on the compact rectangle gives the uniform first-order expansion
where . One can obtain this expansion by integrating the derivative along the state-parameter segment. The common rectangle is convex in those variables, and all the segments remain inside it.
The linear equations for have unique solutions on this short interval: their integral maps are contractions on complete spaces of continuous matrices, with factor at most . Let . Subtract its integral equation from that for . The remainder has zero initial value and satisfies
Gronwall yields . This proves the derivative as a linear approximation to the actual solution map. It proves more than the existence of separate directional derivatives.
The coefficients depend uniformly continuously on nearby data along this interval, because the curves converge uniformly and the derivatives of are uniformly continuous on the compact rectangle. For two sets of coefficients and solutions, write and similarly for the other quantities. Comparing the linear integral equations and applying Gronwall gives
The sizes of the matrix solutions are bounded by and . Thus depend continuously on the data. Finally is jointly continuous. These facts give the stated joint regularity.
To reach a longer compact reference interval, cover its trajectory by finitely many local rectangles. Subdivide time so that the reference curve travels through them in order. At each subdivision point, continuity lets the initial-data neighborhood be reduced so that nearby curves enter the next rectangle. Only finitely many reductions are needed; the comparison constants are taken on these finitely many compact rectangles, and local uniqueness makes the successive solutions agree on overlaps. The longer solution map is a finite composition of the local maps. This proves dependence on any fixed compact reference interval, while making no assertion at a moving maximal endpoint.
Proof X-ray
Differentiating the equation is the candidate step, not the justification
The formal differentiated ODE suggests and . The uniform Taylor remainder and Gronwall estimate prove that those candidates approximate actual changes in the solution. The common existence interval is needed before either comparison makes sense. Losing any one of these steps leaves a gap between a plausible sensitivity equation and a differentiability theorem.
Why compact containment permits continuation
Proof
(Proof of the continuation criterion)
Local uniqueness makes any two solutions through the initial data agree on their common interval. Taking their union gives a well-defined solution on the union interval, and this is the unique maximal solution.
Now suppose its graph for lies in a compact set . Continuity bounds there by some . Consequently on that tail. The curve is Cauchy as , so it has a finite limit . Since is closed, . Local existence from that point supplies a solution past . The integral equation holds up to by continuity, and the two pieces join with derivative . They form an extension, contradicting maximality.
An important consequence is a sufficient global-existence criterion. If and, for fixed ,
where are continuous nonnegative functions on , no finite endpoint can occur inside . On any finite closed time interval, bound them by . The integral equation and Gronwall bound the state by . A hypothetical finite endpoint would therefore have a tail graph in a compact time interval times a closed state ball, allowing continuation. The backward argument is the same after time reversal. This growth condition is sufficient, not necessary.
Proof
(Proof of the event-time formula)
Finally set . The differentiability theorem makes a scalar function near the reference event, with and . Apply the scalar implicit theorem to . The denominator is the speed of change of the constraint along the trajectory. If it vanishes, a trajectory can touch the constraint without crossing it, and a differentiable root time need not exist.
Worked examples
Worked example 3.2.8
(Sensitivity of a branch without an explicit solution)
For , the point solves the constraint and there. The implicit theorem gives a unique local branch. Differentiation gives
The branch is smooth because is smooth. Differentiating once more gives , hence . Thus . A local response can be computed even though a useful exact solution formula is unavailable.
Worked example 3.2.9
(A growth rate changes both the trajectory and its crossing time)
For , , the solution is . The variational equations give and . To reach a fixed level , take . The crossing time is . For , the event denominator is , and
As , this time diverges although the flow varies smoothly on every fixed finite interval. Smooth dependence on compact time intervals does not promise an event at a uniformly bounded time.
Worked example 3.2.10
(A bounded state can reach an excluded boundary)
For on and , integration gives . Its maximal interval is . At the finite right endpoint the state tends to zero, not infinity. There is no extension as a solution taking values in . The graph fails to stay in a compact subset of , exactly as the continuation criterion allows. Its initial-state sensitivity also becomes unbounded near that endpoint, even though it is finite and continuous on every compact subinterval.
Counterexample mode
A smooth flow can have a nonsmooth event time
Let , , so . The constraint gives . At the event derivative is . For positive there are two nearby roots ; for negative there are none. Even choosing only the nonnegative root does not give a differentiable branch at zero. Smoothness of the flow does not replace the transverse-event hypothesis.
Common mistakes
Common mistake
(Different initial values for the two sensitivities)
The two sensitivity matrices have different initial conditions. Changing changes the curve at the initial time, giving ; changing while keeping fixed gives . The inhomogeneous term records the continuing effect of the parameter on velocity. If the initial state itself is , the total derivative is instead.
Common mistake
(Bounded or unique does not mean global)
Boundedness of the state prevents finite escape only when it also keeps the graph away from the boundary of the time-state domain. Nor does local uniqueness imply global existence.
Common mistake
(A local event root need not be the first event)
At an event, the implicit theorem selects a root near the reference root. To call it the first event, one must separately rule out earlier zeros, for example by a strict sign condition on the preceding part of the reference trajectory and suitable uniform control.
Summary
An invertible constraint derivative turns data changes into a local solution branch. A common ODE existence interval and Gronwall control turn formal sensitivities into true derivatives. Compact containment prevents a maximal solution from stopping at an interior time. Combining the solution map with a transverse scalar constraint then gives a differentiable local event time. Each conclusion has its own domain and nondegeneracy conditions.
Exercises
Quick check
(For near , prove there is a smooth local branch and compute its first two derivatives at zero.)
Solution · Guided solution
, which equals one at the base point. The implicit theorem applies, and smoothness follows by repeated differentiation. Differentiating gives , hence . A second differentiation gives , so . The resulting expansion is .
Quick check
(Solve , . Compute and verify both variational equations. Interpret their limits as .)
Solution · Guided solution
The solution is . Thus and . Since and , these satisfy , , and , . As time increases, sensitivity to the initial state tends to zero while sensitivity to the equilibrium parameter tends to one. These are different responses even in this elementary scalar equation.
Quick check
(On a common forward interval with , suppose state and parameter Lipschitz bounds are , . If initial states differ by at most and parameters by at most , give the uniform bound proved above.)
Solution · Guided solution
The bound is . It is a guaranteed upper bound, not a claim that the actual error reaches this size. Its validity depends on both trajectories remaining in the common rectangle where the quoted constants hold.
Quick check
(For with fixed real and arbitrary real initial state, prove the maximal solution is defined for all real times. Give a simple bound on its displacement from the initial value.)
Solution · Guided solution
The vector field is smooth on all of , and . The integral equation yields on the existence interval. A finite endpoint would therefore have a tail graph in a compact rectangle of , and the continuation theorem would extend the solution. Neither endpoint can be finite, so the solution is global in both directions.
Quick check
(Let , , and consider near . Determine the local event time and its derivatives with respect to and . Explain why the other root does not violate local uniqueness.)
Solution · Guided solution
Here , so . Near the reference point the correct branch is . Thus and , equal to at the reference point. The denominator is nonzero there. The second root lies near , outside the chosen neighborhood of the reference time one, so it does not contradict local uniqueness.