An unknown slope, not an unknown endpoint
Suppose a scalar position satisfies
An initial-value solver needs both and . Shooting supplies a trial slope and compares the endpoint of the resulting trajectory with .
Definition 9.2.1
(Shot, endpoint map and shooting residual)
For a slope , the shot is the solution of with , . When exists on all of , the endpoint map is and the shooting residual is .
A zero of is a boundary solution only when the trajectory actually exists up to . In a nonlinear problem, trajectories cannot generally be combined by superposition, so is a genuinely nonlinear scalar function.
The endpoint map also describes sensitivity. If changing barely moves , a small endpoint error can require a large slope correction. We will derive an equation for this response before proposing an algorithm.
A condition that makes the shot unique
Theorem 9.2.2
(Monotone nonlinear shooting)
Let and be jointly continuous on and suppose for a finite constant . Then for every real , problem (1) has exactly one solution. For every slope , the endpoint map is continuously differentiable and
If are two slopes, then . This is a continuous, exact-arithmetic statement, not a bound on a numerical integrator's error.
The proof rests on one comparison fact about a linear equation with a nonnegative coefficient.
Lemma 9.2.3
(A nonnegative coefficient keeps the response growing)
Let be continuous and nonnegative on , and let , , . Then and for all .
Proof
(No first zero can occur)
Since , the function is positive immediately to the right of . If it had a first zero , then on , so there, hence and , a contradiction. Thus on , so on the whole interval, whence and .
Proof
(Monotone shooting from the variational equation)
The upper bound on makes globally Lipschitz in . Together with continuity on the finite interval it guarantees that each initial-value trajectory exists uniquely through , so is defined on all of . Continuous dependence and differentiation of the initial-value problem give , satisfying
This is the variational equation: it is linear even when the original equation is nonlinear, but its coefficient depends on the current shot. By the lemma, , so .
For , integration gives . Thus is strictly increasing and tends to and as does. The intermediate value theorem supplies exactly one slope with . The same inequality gives the stated sensitivity estimate. Every solution of (1) has some initial slope, so this also proves uniqueness of the boundary solution.
Proof X-ray
Why the sign matters
Each hypothesis has one job. The upper bound is used only to make every shot exist on all of , so that is defined for every real ; without it the domain of can have gaps, as the blow-up example below shows. Continuity of is used only to differentiate the shot and obtain (3). The lower bound is used only in the lemma, to stop the sensitivity from returning to zero. The key move is that a uniform lower bound on , integrated, does two things at once: it makes injective, which gives uniqueness, and it makes unbounded in both directions, which together with continuity gives existence through the intermediate value theorem. The final step needs the observation that every boundary solution is some shot; without it, uniqueness of the slope would not imply uniqueness of the solution.
Newton's correction and numerical trajectories
When the hypotheses hold, compute and by integrating (1)'s initial-value equation and (3) together. Newton's update is
The exact denominator is at least . This makes the correction locally meaningful but does not guarantee that an arbitrary first guess converges. A bracket with permits bisection, or a safeguarded Newton step that stays inside the bracket. These statements concern the exact endpoint function; numerical signs near zero may be wrong unless the IVP error is controlled.
In code, one instead has a numerical endpoint and should differentiate the discrete update to obtain . Using the continuous sensitivity equation with the same differentiable one-step scheme gives this derivative for Euler, as the next example shows. A root of matches the boundary of the computed trajectory, not necessarily that of the true trajectory. Refine the mesh or estimate IVP error before interpreting a small numerical residual as accuracy.
Worked example 9.2.4
(A nonlinear endpoint computed with Euler)
Take , , , , and . Here , so the theorem applies. With and , explicit Euler gives
Starting from , the positions are , , ; also . Therefore
At , and , so one discrete Newton step gives . Its residual is : one step has not exactly solved even the discrete problem. The next update is
This calculation is a checkable three-step numerical model, not the exact continuous shooting slope. Its discretization error remains to be assessed.
Worked example 9.2.5
(Bisection on the same discrete endpoint)
The discrete map inherits the theorem's monotonicity: , so has exactly one root. The slopes and bracket it, since and . Bisection evaluates , giving the bracket , then , giving , then , giving . After three evaluations the root is known to within an interval of length ; each further evaluation halves it. The Newton iterates above were already within of the root after two steps, but bisection is guaranteed by the bracket alone, without any assumption about the first guess.
When the hypothesis fails
The sign condition cannot simply be dropped. The linear equation already shows what goes wrong.
Counterexample mode
Well-posed shots, ill-posed boundary problem
Claim: if every initial-value problem , , has a unique global solution, then the boundary problem (1) has exactly one solution. Take on , . The trajectory is , so for every slope. If there are infinitely many boundary solutions; otherwise there are none. Here , the upper bound holds, and every shot is perfectly well posed. The sensitivity vanishes at the endpoint, which is exactly what the lemma excludes when . The repaired statement is the theorem: global well-posedness of the shots together with .
Conversely, a failed hypothesis does not prove failure: may be negative in a problem whose endpoint response is nevertheless nonzero. The theorem is sufficient, not necessary.
Worked example 9.2.6
(A shot that never reaches the endpoint)
For , the function satisfies and for , and it blows up at . Take and . The shot with slope has initial data , , which are those of this function with ; since is locally Lipschitz, the shot is that function. It exists on but not at , so is undefined. The hypothesis that fails is the upper bound: is unbounded. A root search must account for trajectories that cease to exist before .
Common mistake
(Combining two shots by superposition)
For a linear equation, a combination of two shots is again a solution, which is why linear shooting needs only two integrations. For , if are solutions then has , which differs from wherever the shots differ. Interpolating between shots may suggest a better slope, as in the secant method, but the combined trajectory is not a solution; the new slope must be integrated again.
Common mistake
(Freezing the variational coefficient)
The coefficient in (3) belongs to the current shot. Reusing the coefficient of an earlier shot gives a different derivative. In the Euler example, the trajectory with is identically zero, so its coefficient is and its discrete sensitivity is , whereas at the correct value is . Iteration (4) with a frozen denominator becomes a chord iteration: it has the same fixed point but in general converges only linearly instead of quadratically.
Multiple shooting, general , stiff IVP solvers, certified numerical brackets and global Newton convergence remain outside the unit.
Practice
Quick check
(On , solve , , by shooting. Compute and one Newton step from any .)
Solution · Answer
, so , and . Formula (4) reaches in one step from any because is affine. The boundary solution is .
Quick check
(Under the theorem, two shots have endpoint values differing by on an interval of length . Bound the difference of their initial slopes. Is the bound a numerical error certificate by itself?)
Solution · Answer
The slope difference is at most . This compares exact endpoints and slopes; approximate endpoints carry additional errors that must be bounded separately.
Quick check
(For on , , find the endpoint map. Explain why this does not contradict the theorem.)
Solution · Answer
and , so each endpoint has a unique shot. The sign condition fails, but it was never claimed necessary.
Quick check
(For the Euler example, differentiate the Euler updates with respect to and verify the first two sensitivity states , where . What initial values are required?)
Solution · Answer
Start from . Differentiate to get and . Thus and, since , , which equals at . One more step gives . These are derivatives of the discrete trajectory, not automatically exact continuous sensitivities.