From a single trajectory to a family
An initial value problem produces a trajectory, but many questions compare nearby trajectories. If we alter the starting state by a small vector, how much does the state at a fixed later time change? If a coefficient is measured imperfectly, what is the first-order effect on the prediction? The answers are derivatives of the solution with respect to data. They are not derivatives with respect to time, though they satisfy another differential equation in time.
There is a preliminary issue: nearby solutions must exist at the time being compared. For , , the solution is only until . A formula for sensitivity at time is meaningful only for data whose solutions actually reach . We first secure a common finite interval, then differentiate.
The theorem and the two sensitivity equations
Throughout, , and are open, and and its partial derivatives are continuous. Fix , , and a solution of
on a compact interval , with graph inside . For everything below holds after reversing time.
Definition 1.10.1
(Variational equation and sensitivity matrices)
Along the reference solution put
The variational equation in the direction is the linear initial value problem
The state sensitivity is the matrix solving , , and the parameter sensitivity is the matrix solving , .
Both coefficient matrices are evaluated along the whole reference trajectory, so they generally depend on even when does not. Equation (1) is linear with continuous coefficients, so its solution exists on all of ; the issue is whether it describes the nonlinear problem.
Theorem 1.10.2
(Differentiable dependence on initial data and parameters)
Under the standing hypotheses, some neighborhood of has the property that every gives a unique solution of , , defined on all of . The map is continuously differentiable as a map into with the supremum norm. Its derivative at in the direction is the solution of (1). Consequently
where and , and
uniformly for .
The zero initial value for assumes that the starting state is held fixed as changes. If , the chain rule replaces it by .
Why nearby solutions reach the endpoint
The first assertion of the theorem, existence on the whole of , is the part that fails without care, so we isolate it.
Lemma 1.10.3
(Nearby solutions stay in a tube on the whole interval)
There are , a neighborhood of and a constant such that for the solution exists on , satisfies there, and .
Proof
(Tube, Grönwall estimate and first exit)
The reference graph over is compact and lies in the open domain. Consequently a sufficiently thin closed tube of radius around it, together with a small closed parameter ball about , lies in the domain. On this compact set is bounded, and has a finite bound . Each state cross-section of the tube is a ball, so the mean-value inequality gives the common state-Lipschitz bound there.
While a nearby solution remains in the tube, subtract the integral equations. Uniform continuity in the parameter gives a number such that, for ,
Grönwall's inequality bounds the distance by . Choose the initial and parameter differences so this is less than , half the tube radius. If the solution left the tube before , there would be a first time at which its distance from equals ; the estimate holds up to that time and gives a distance below , a contradiction. Local existence and uniqueness restart the solution as needed, and the compact continuation theorem prevents a finite endpoint before while the graph stays inside this compact tube. Since is continuous, it is bounded by some on the compact set, so by the mean-value inequality in the parameter ball; this gives the linear bound with when pairs carry the norm . This is a local statement around one reference trajectory on one fixed interval; it is not a global existence theorem.
Proof of differentiability
Proof
(Linearizing the augmented equation)
Combine state and parameter into , with augmented equation . It is continuously differentiable in . Let be the reference solution and let start at . For small , the lemma gives for a fixed constant ; bounded on the compact tube gives this linear bound in the parameter difference.
Let solve , . Uniform continuity of on a compact neighborhood of the reference graph gives
Subtracting the two integral equations and applying Grönwall yields . Because the linear solution depends linearly and continuously on , this proves the Fréchet derivative in the supremum norm, rather than merely formal differentiation along coordinate directions. The parameter component of is constant and equal to ; its state component is exactly (1).
For continuity of the derivative, let the base data vary. The corresponding trajectories converge uniformly on and hence the coefficient matrices along them converge uniformly. Subtracting their linear integral equations and applying Grönwall again shows that their derivative operators converge in norm. This completes the proof.
Proof X-ray
Three distinct uses of compactness
The proof has two layers, and each hypothesis enters at a definite place. Compactness of the reference graph is used three times: to fit a closed tube inside the open domain; to bound on that tube, which turns into the Lipschitz constant of the Grönwall step and keeps nearby solutions from escaping; and to make uniformly continuous there, which is what makes the Taylor remainder small uniformly in rather than only pointwise.
The key move is the augmentation : it turns a parameter into an initial value with zero velocity, so one argument handles both kinds of data, and the state component of the linear solution is the variational equation (1). The argument closes with the same Grönwall inequality used twice, once on the nonlinear difference and once on the difference between the nonlinear perturbation and its linear prediction. Without continuity of the second use fails, as the counterexample below shows.
The variational equation also has a description that does not mention perturbations at all.
Concept lensStructural
The sensitivity matrix as the derivative of the flow map
For a field without parameters write for the state at time of the solution starting at at time . Uniqueness gives the composition rule , and the theorem says . The chain rule then gives : sensitivity over a long interval is the product of sensitivities over its pieces. In this description is the principal fundamental matrix of the linear system , and each of its columns is the first-order image of one coordinate perturbation of the initial state.
Worked examples
Worked example 1.10.4
(Logistic sensitivities, checked against the explicit solution)
Consider , with . Separation and direct verification give, for all real ,
Write and . Then , and, using , , which equals because .
At , and . Here and , so the variational equations are and ; differentiating the displayed formulas verifies both. For positive time, : increasing the growth parameter increases the state to first order. This conclusion depends on the given initial state in and on ; it is not an assertion about every logistic initial value.
Worked example 1.10.5
(Linearizing about an equilibrium)
Take the same equation with , so the reference solution is the equilibrium for every . Then is constant and . The sensitivity equations become
Both agree with the explicit formulas at : and . The interpretation is consistent: for a perturbation of the initial state decays like , and changing does not move the equilibrium, so the first-order parameter effect is zero. In this edge case the theorem reduces to the familiar linearization at a constant solution.
Worked example 1.10.6
(Blow-up limits the interval of comparison)
For , , the solution is for . Differentiating in ,
Check: , and with . For a prescribed , the data whose solutions reach are exactly . For each such the theorem applies on , but tends to as from below. The neighborhood in the theorem must shrink as the reference datum approaches the boundary of this set; the theorem is local and gives no common interval across it.
Counterexample mode
A Lipschitz field need not give a differentiable flow
Claim: if solutions exist, are unique and depend continuously on the initial state, they depend differentiably on it. For , , the field is globally Lipschitz and every solution exists for all time. At any fixed , the solution map is for (since then and ) and for (since then and ). Its one-sided derivatives at zero are and , which differ. Existence, uniqueness and continuous dependence therefore do not by themselves imply differentiable dependence; the hypothesis that fails is continuity of , since is not differentiable at zero.
Common mistakes
Common mistake
(Freezing the coefficient at the initial point)
It is tempting to replace by the constant and write . The coefficient must be evaluated along the reference trajectory. In the logistic example, changes sign as passes , and the constant-coefficient guess is not the sensitivity unless is an equilibrium.
Common mistake
(Using the zero initial value when the start depends on the parameter)
The condition is correct only when the initial state is held fixed while varies. If the problem prescribes , the total parameter sensitivity solves the same equation but with ; equivalently it equals plus the fixed-start .
Practice
Quick check
(For , , find and and verify their initial conditions and equations.)
Here and .
Solution · Answer
The solution is , so and . They satisfy , , and , , as (2) requires.
Quick check
(For , , solve the sensitivity equations (2) without first solving for , and then check against the explicit solution.)
The coefficients are and .
Solution · Answer
From , we get ; from , we get . The explicit solution is , whose partial derivatives are and , in agreement.
Quick check
(Suppose the logistic example starts instead at for . What is the initial value of the total parameter sensitivity, and why is it not zero?)
Apply the chain rule at .
Solution · Answer
The initial value is . Changing the parameter also changes the prescribed initial state, so the total derivative includes that initial variation as well as the forcing term .