A stable homogeneous equation forgets its initial state. A forced equation continually receives new disturbances, so forgetting its past does not imply approaching zero. Variation of constants separates these two effects: one term propagates the initial state, while an integral propagates each disturbance from the time it enters.
For a nonlinear equation, the remainder beyond the Jacobian acts as another forcing term, but one depending on the unknown trajectory. This dependence creates a feedback estimate. A strictly decaying linear part can absorb sufficiently small nonlinear feedback. The proof must also keep the trajectory in the neighborhood where that estimate is valid.
Motivation
The eigenvalues of a constant matrix describe long-time homogeneous behavior. They do not alone specify the size of a transient excursion or the response to sustained forcing. A matrix can have only negative eigenvalues and still amplify an initial disturbance before it decays. In a nonlinear problem, that excursion may reach a region where the linear approximation no longer controls the dynamics.
The useful information is therefore an estimate with both a decay rate and a prefactor. The rate measures how quickly the linear propagator loses memory; the prefactor records possible amplification. Both enter the permissible nonlinear size and the forcing bound. Retaining these constants makes the indirect stability method quantitative.
Definitions
Fix a vector norm and its induced matrix norm. For a continuous matrix on , let , , and define the evolution matrix
Linear initial-value uniqueness makes invertible. The two-time notation matters: estimates beginning at different input times require control of , not only of .
Definition 3.2.1
(Uniform exponential propagation)
The linear equation has a uniform exponential estimate if some and satisfy for every . For a constant matrix, . A constant matrix is Hurwitz if every eigenvalue has negative real part.
For the nonlinear part, let be open and contain zero. Assume , , and write
The forced equation is , where the input is continuous on . Local existence and uniqueness follow from continuity in time and local Lipschitz regularity in the state. A bounded-input assumption below means ; it does not mean that tends to zero.
Theorem/proposition
Theorem 3.2.2
(Variation of constants and a linear forcing estimate)
For continuous , the solution of is
If the evolution matrix has the uniform exponential estimate, then
For bounded input the last term is at most . If the bounded continuous input tends to zero, the solution tends to zero.
Lemma 3.2.3
(An exponentially weighted comparison lemma)
Let be nonnegative continuous functions on and suppose
Then
Theorem 3.2.4
(Local nonlinear forcing bound with continuation)
Suppose is constant with for . Choose with , and such that throughout that ball. Put . If is continuous, bounded by , and
the solution exists for every forward time, remains in , and satisfies
Consequently . If the input tends to zero, so does the state.
Theorem 3.2.5
(Lyapunov's indirect method as a quantitative corollary)
If near zero, , and is Hurwitz, then zero is locally exponentially stable for the unforced equation. More precisely, choose any valid linear estimate . A sufficiently small radius permits , giving
The forced conclusion is a local robustness estimate, not asymptotic stability of zero under every small input. With a nonzero constant input, zero may cease to be an equilibrium at all.
Proof sketch or proof idea
Propagate disturbances from their actual entry times
Proof
(Variation of constants and the forcing bound)
Set . The homogeneous terms cancel, leaving . Integration and multiplication by give the variation formula. Conversely, differentiating that formula recovers the equation and initial value. Continuous coefficients ensure linear solutions extend across every finite time interval; an integral Gronwall bound prevents finite escape there.
Apply the evolution estimate inside the integral and use the triangle inequality. For an input bounded by , integrating the exponential kernel gives . To prove the vanishing-input assertion, split the convolution at a fixed time . The contribution from decays exponentially as . If for , the remaining contribution is at most before the prefactor . Let . Boundedness alone would not justify this last step.
Concept lensAlgebraic
The forced solution as a superposition of launched solutions
For each input time , let solve the homogeneous equation on with , so . The variation formula reads
the forced solution is a continuous superposition of homogeneous solutions, one launched at every input time with the input value as its initial state. The forcing estimate says precisely that each launched solution decays at a uniform rate measured from its own launch time.
Proof
(An integrating factor for the comparison lemma)
For the comparison lemma, multiply by and set . Then
Since , . Multiply by , integrate and then divide by . This yields the asserted kernel with rate . The lemma is valid even if that rate is nonpositive; a positive rate is the additional condition needed for decay and a finite bounded-input gain.
Absorb the remainder, then justify staying local
Proof
(Absorb the remainder, exclude a first exit, continue)
On any interval where the nonlinear solution stays in , variation of constants gives
Taking norms gives the comparison lemma with , , . Hence the nonlinear estimate holds on that interval. In particular the state norm is at most
Suppose there were a first exit through the sphere of radius . The estimate holds up to that time by continuity and gives a norm at most there, a contradiction. The solution therefore stays in the smaller closed ball for as long as it exists.
If its maximal forward time were finite, continuity would bound on that compact ball and bound on the finite time interval. Thus would be bounded, making Cauchy at the endpoint. Its limit lies inside , and local existence at the limit extends the solution. This contradiction proves forward existence for all time. The convolution estimate and the vanishing-input argument now apply on arbitrarily long intervals, establishing the remaining conclusions.
Proof X-ray
The first-exit argument closes a conditional estimate
The estimate is proved in the wrong order on purpose. First it is derived conditionally, on an interval where the trajectory is assumed to stay in , because only there is known. The hypothesis is used inside that derivation, to give the forcing integral the finite gain . The strict inequality is then used once, at a hypothetical first contact with the sphere of radius : the conditional estimate still holds up to that moment and places the state at norm at most , which is not on the sphere. With the trajectory confined to the compact ball , boundedness of there and of on finite intervals excludes a finite maximal time. Only now is the conditional estimate valid for every . If were allowed, the first-contact step would give no contradiction.
Proof
(The indirect method from the forcing bound)
For a Hurwitz constant matrix, a Jordan block contributes a polynomial in times . Choose strictly smaller than the absolute value of the largest real part among the eigenvalues. Each polynomial times the remaining decaying exponential is bounded. Combining finitely many blocks and the change-of-basis norms gives . A nontrivial Jordan block generally prevents taking the spectral decay rate itself with a finite prefactor.
Since , differentiability gives . Shrink until the desired works. The nonlinear theorem with then proves the indirect method, including continuation and an explicit exponential bound. This argument explains which part of the conclusion is local: the small remainder estimate is only available near the equilibrium.
The same forcing distinction appears in a quadratic Lyapunov proof
For this paragraph use the Euclidean norm. The matrix
is symmetric positive definite and satisfies . Exponential decay gives convergence; for . Integrating the derivative of the matrix integrand proves the identity. Put and . Along the forced trajectory,
If the local remainder bound also has , then
The middle step is Young's inequality, obtained by expanding a nonnegative square. On a solution already kept in the valid region, integration gives
This energy estimate provides a second way to measure a sustained response. It also shows why a quadratic proof cannot simply discard the forcing term and retain strict decrease everywhere. With , the positive contribution may dominate near the origin.
Worked examples
Worked example 3.2.6
(A decaying memory with a persistent periodic response)
For , , variation of constants gives
Differentiation verifies the forcing and initial value. The homogeneous contribution decays, but the remaining periodic response has amplitude , so the state need not tend to zero. The linear estimate with , gives . Its ultimate bound is valid but conservative compared with the exact amplitude.
Worked example 3.2.7
(Negative eigenvalues can hide transient amplification)
Use the infinity norm and . Then
Since , one valid estimate has and . Starting from , the first coordinate at time one is , whose magnitude exceeds the initial infinity norm when . The matrix is still exponentially stable. With constant input , its equilibrium is , so the spectral real parts alone do not determine the forcing gain either.
Worked example 3.2.8
(A nonlinear radius and a certified forcing allowance)
For , take . On , . With this gives . If and , then
The solution is therefore global forward in the small region, with . For the constant input , the small equilibrium is . Its derivative is negative. The forcing shifts the equilibrium; the estimate never promised convergence to the unforced origin.
Counterexample mode
Pointwise negative time-dependent coefficients do not give the needed kernel bound
For , . Every fixed-start homogeneous solution tends to zero, and the scalar coefficient is negative at every time. There is nevertheless no uniform exponential estimate: setting would require for all , which is impossible for positive .
Add an arbitrarily small constant input . Direct integration gives
which grows like . The missing hypothesis is uniform decay of the two-time evolution matrix. Instantaneous eigenvalue signs and convergence for each homogeneous initial state do not replace it.
Common mistakes
Common mistake
(A forced state need not tend to zero)
Local exponential stability concerns the unforced equilibrium. Under a persistent input, a bounded ultimate response is the appropriate conclusion here. Even tends to , not zero. If the input vanishes, use the convolution argument rather than replacing its whole history by one supremum bound, which loses the decay information.
Common mistake
(Dropping the prefactor or the region)
The prefactor is also not automatically one. Ignoring transient amplification can invalidate a proposed nonlinear trapping radius. Similarly, a remainder is small relative to only within a specified region. For , zero is locally exponentially stable, but initial values greater than one escape in finite forward time. The local theorem makes no claim about them.
Common mistake
(Controlling only the propagator from time zero)
Finally, an estimate for alone need not control disturbances entering at arbitrary times. The linear forcing theorem assumes the same constants for every . A nonautonomous equation also need not have when the matrices at different times fail to commute; the fundamental-matrix construction avoids that invalid simplification.
Summary
Variation of constants separates propagated initial data from propagated input. A uniform exponential kernel gives a finite forcing gain. A small nonlinear remainder consumes part of the decay margin, leaving rate . A first-exit and continuation argument makes the conditional local estimate valid for all forward time. This proves the indirect method and distinguishes persistent bounded response from genuine convergence to zero.
Exercises
Quick check
(Solve , . Show directly that the forcing contribution tends to zero and compare it with the bounded-input estimate.)
Solution · Answer
The solution is . Both terms tend to zero. Using only would bound the input response by , whose limit is nonzero. The exact convolution or the vanishing-input theorem retains the information lost by that coarse supremum estimate.
Quick check
(For , compute the matrix satisfying . For on the Euclidean ball of radius , verify the smallness condition used in the energy estimate and give its resulting differential inequality.)
Solution · Answer
, so . On the stated ball, , and . Thus, while the trajectory remains in this ball, the proved energy estimate gives , where . The region qualification must remain unless a trapping argument has also been supplied.
Quick check
(Suppose , , , , and . Check the nonlinear forcing certificate and write the resulting time-dependent bound.)
Solution · Answer
, and . The theorem applies and gives . The sharper final bound comes from retaining the exponential factor in the forcing integral; the simpler strict-margin test was deliberately conservative.
Quick check
(Solve for and find its finite forward escape time. Explain why this does not contradict the indirect method at zero.)
Solution · Answer
The substitution gives , hence . For its denominator first vanishes at , and as . The derivative of the vector field at zero is , so the indirect method controls sufficiently small initial states, not these data beyond the other equilibrium one.
Quick check
(Retain the nonlinear hypotheses and , but assume only that is continuous and . Prove forward existence and convergence to zero if . Do not assume the input is uniformly bounded on the whole half-line.)
Solution · Answer
On each finite interval, continuity bounds the input and the same variation/comparison argument applies. Since , the state is bounded by the strict radius . The first-exit argument and compact continuation work as before, using boundedness of on each finite interval only. For convergence, split the convolution at . The earlier contribution decays exponentially; the later contribution is at most , which tends to zero as . Thus the state tends to zero even without a global input supremum.