From a boundary condition to an unknown slope
An initial-value method expects both position and velocity at the same endpoint. A two-point boundary problem instead specifies one value at each end. Shooting guesses the missing initial slope, integrates an initial-value problem, and adjusts that slope until the far endpoint matches. The adjustment is an additional numerical problem: even a highly accurate trajectory need not produce a stable slope if the endpoint hardly changes when the slope changes.
Consider
where and are continuous real functions. For each initial slope , standard linear initial-value theory gives a unique solution.
Definition 9.1.1
(Shooting trajectories and the endpoint denominator)
Let solve the equation with the original forcing and initial data , . Let solve the corresponding homogeneous equation with , . The number is the endpoint denominator: it measures how strongly the far boundary value responds to the initial slope.
Linearity gives for the solution with initial slope : the difference solves the homogeneous equation with initial data , as does . The endpoint equation is therefore
Proposition 9.1.2
(The linear shooting alternative)
If , the boundary-value problem has exactly one solution, with initial slope . If , it has no solution when and infinitely many when .
Proof
(Every solution is a shooting trajectory)
Any solution of the boundary problem has and some slope , so by uniqueness for initial-value problems . Conversely satisfies the boundary conditions exactly when solves (1). Equation (1) is a linear equation in one unknown. If it has the single root displayed. If it reads : it has no root when , and every is a root when . Distinct slopes give distinct solutions, because their initial derivatives differ.
Thus existence of every initial-value trajectory does not establish uniqueness of the boundary problem. The difference of two candidate trajectories is , whose far value is : when the boundary cannot distinguish the slopes, and when is small but nonzero it distinguishes them only weakly.
Concept lensStructural
The denominator detects a homogeneous boundary solution
The condition says exactly that the homogeneous problem , , has only the zero solution. Indeed itself is such a solution when ; conversely a nonzero solution has , so by initial-value uniqueness, and then . For on this is the statement that is not one of the Dirichlet eigenvalues .
A concrete numerical algorithm
Put , and . On , , advance two vectors by explicit Euler:
The subscript 1 below denotes the first component. If , set
Then and exactly in exact arithmetic. The latter is a boundary match for the discrete trajectory, not proof that its interior is accurate. The same superposition idea works with any linear IVP discretization, but the estimate below is specifically for Euler on a uniform grid.
Theorem 9.1.3
(Euler shooting converges when the endpoint map is nonsingular)
Assume are continuously differentiable on and . For all sufficiently large , formula (2) is defined and
where depends on the fixed problem and can grow as decreases. The norm can be any fixed norm on . This is a finite-interval, exact-arithmetic statement; it does not bound roundoff or assert a constant uniform over nearly singular problems.
Proof: residual, propagation and division
Proof
(Residual, propagation and division)
Write and . Their derivatives are continuously differentiable and their second derivatives are bounded because the coefficients are . Hence Taylor's theorem gives, for either exact vector, a one-step defect bounded by in a fixed norm. Since is bounded, the error of its Euler approximation satisfies
Summing the geometric weights gives , including the case . Enlarge the constants if necessary to bound both the vector errors and their first components (coordinate projection need not have norm one for an arbitrary chosen norm). Consequently, for constants independent of ,
For sufficiently small , , so the discrete denominator has magnitude at least . Subtracting the exact endpoint relation yields
Thus . The numerical are uniformly bounded, so
This proves both the slope and trajectory estimates, since .
Proof X-ray
Why the denominator matters
The proof has three stages, and only the last uses . Stages one and two are standard Euler theory applied separately to the forced trajectory and to the slope response: a local defect of size , amplified by at most over steps, gives first-order errors in and . These hold for every problem with coefficients, singular or not. The third stage divides by the computed denominator . The key move is to wait until is small enough that is within of ; this keeps the discrete denominator away from zero and costs a factor in the slope error. That factor then multiplies the bounded in the trajectory estimate, which is why blows up as . If , the third stage has nothing to stand on: the discrete denominator may still be nonzero, but it tends to zero and the quotient need not converge.
Worked examples
Worked example 9.1.4
(Two Euler steps for a curved solution)
Solve on with , . Here . With , Euler applied to gives
Hence , and the computed positions are . The exact solution is , with initial slope and midpoint . The endpoint is matched exactly, yet the midpoint error is about . A boundary residual alone therefore cannot measure the whole trajectory error.
Worked example 9.1.5
(A forced problem: the slope error is exactly )
Solve on with , . Here and , so , , and . For Euler with , the slope response is reproduced exactly, , so . The forced vector has and
using . Hence and : for the computed slope is , for it is . Halving the step halves the slope error, which is the first-order rate of the theorem.
Worked example 9.1.6
(Sensitivity near resonance)
For on with , the slope response is and , so with . At , and raising by raises the slope by . At , close to , and the same change in moves the slope by about , roughly thirteen times as much. The problem is still uniquely solvable; it is merely ill-conditioned, and any error in or in the computed is amplified by the same factor .
Resonance and sensitivity
For , , , and , the homogeneous slope response is and . At the right endpoint is zero for every initial slope: gives infinitely many solutions, and gives none. Euler may nevertheless produce a small nonzero discrete denominator and a seemingly unique numerical answer. This does not repair the continuous problem.
Counterexample mode
A defined discrete slope without a continuous solution
Claim: if the discrete denominator is nonzero, the computed approximate a solution of the boundary problem. Take , , , which has no solution. Euler gives with . With , and , so and . Refining gives , , for , hence , , . Indeed , so the computed slopes grow without bound. Each discrete problem is solvable, but the theorem's hypothesis fails and there is nothing to converge to. A computed slope that keeps changing under refinement is the practical signal.
If and is fixed, changing only by changes the exact slope by . Near , this amplification can be large. A small denominator should prompt a conditioning analysis, not an automatic declaration that the integrator is faulty. Roundoff and integration errors in both numerator and denominator also affect the quotient; the exact-arithmetic convergence theorem does not quantify them.
Common mistake
(Integrating the forcing in the slope response)
The second trajectory must solve the homogeneous equation. If it carries the forcing as well, superposition fails. For , a function with , , is , and has second derivative , which is not unless . The forcing belongs to only; only records how the solution responds to the slope.
Common mistake
(Treating an exact endpoint match as an error estimate)
Formula (2) makes whatever the step size, so a zero boundary residual says nothing about accuracy. In the first worked example the endpoint is exact while the midpoint is off by about . Accuracy must be judged by refinement, or by the convergence theorem when its hypotheses hold.
Nonlinear shooting, stiff integration, multiple shooting, adaptive error control and a general conditioning theory remain outside this unit.
Practice
Quick check
(For on , , , find , and the shooting slope. Does Euler reproduce this solution at its grid points?)
Solution · Answer
, , and , so . Euler advances a constant velocity exactly, so all its grid values agree with the exact solution.
Quick check
(For , , , , take one Euler step. Compute the discrete denominator. What does this say about a coarse mesh?)
Solution · Answer
, so , whereas the exact denominator is . A nonzero denominator only makes the discrete matching equation solvable; it does not certify a small discretization error.
Quick check
(For the oscillatory example with , calculate and the change in slope when increases by .)
Solution · Answer
, so .
Quick check
(Suppose and . Explain why two different initial slopes produce different solutions with identical boundary values.)
Solution · Answer
Their difference is . It is nonzero because its initial derivative is , but vanishes at both endpoints since .