Why predict and correct?
The trapezoidal rule averages the slopes at the old and new endpoints. Its new slope depends on the unknown new value, so each step normally requires an algebraic solve. If two previous slopes are already available, an Adams–Bashforth prediction supplies a cheap estimate of that endpoint. We can then evaluate the new slope at the prediction and correct once. This saves an implicit solve, but the resulting method is a new explicit two-step method: a single correction does not make the predicted value equal to the solution of the trapezoidal equation.
Definition 4.3.1
(One AB2 prediction and one trapezoidal correction)
Let , with constant , and . From two starting values compute
After the correction, evaluate and store for the next step. Reusing as though it were changes the algorithm. A one-step starter such as explicit midpoint can supply with error or better; taking arbitrarily need not preserve the desired order.
Concept lensAlgorithmic
One sweep of a fixed-point iteration
Write the trapezoidal equation as with . The corrected value is exactly : one fixed-point sweep started from the AB2 guess. Since has Lipschitz constant , the corrected value is closer to the trapezoidal solution by that factor: . In the step example below, and miss by and , the ratio .
Exact-start accuracy and convergence
Theorem 4.3.2
(Second-order convergence on a finite interval)
Let be globally Lipschitz in its state argument with constant , uniformly in time, and let the exact solution have bounded third derivative on . If and are , then, at grid points , the method satisfies as . The same argument applies in any fixed finite-dimensional norm.
Proof
(Exact-start defect, then a two-level Gronwall argument)
Insert the exact previous values in the predictor. Taylor expansion of around and of shows that the predicted endpoint differs from by uniformly. The exact trapezoidal quadrature of has state defect because is bounded. Replacing the exact endpoint slope by changes the corrected value by at most . Thus the corrected exact-start defect is .
Now compare two numerical histories. Let be the maximum of the magnitudes of their differences at and . Lipschitz continuity bounds their prediction difference by . The correction difference is therefore bounded by . For this is at most with . Apply this to a numerical history and the exact samples, adding the defect at each step. The resulting maximum error obeys
Iterating for at most steps gives , proving the claim. The constants can be enlarged between lines. This is a finite-time result; it does not say that every large step is stable.
Proof X-ray
Why one inexact correction costs nothing in order
The consistency half of the proof turns on a single multiplication. The predictor, started from exact history, misses by . That error enters the corrector only through the slope , which is multiplied by and changed by at most times the state error. The predictor therefore contributes , one order below the trapezoidal defect . This is the key move: a predictor of one order lower accuracy than the corrector is enough.
The stability half has to handle a two-step method, so it tracks , the larger error on two consecutive levels. The Lipschitz constant is used twice, once through the predictor, where the three-halves and one-half weights give , and once through the corrector. The step bound absorbs into . The starting hypothesis enters only at the end, where the discrete Gronwall bound multiplies it by . Since can be very large for a stiff problem, this finite-time theorem says nothing about large steps; the linear stability interval below addresses that question.
The stability of the algorithm actually run
Set and . Substitution eliminates the predictor:
Each characteristic root is a recurrence mode. Following the bounded-root convention, a value of is stable when both roots have modulus at most one and roots on the unit circle are simple. A repeated unit root is excluded because it also produces .
Theorem 4.3.3
(Negative-real stability interval)
For , , the one-correction algorithm is stable exactly for . For both roots lie strictly inside the unit disk. At the double root makes the recurrence unstable.
Proof
(A root-location test for a real quadratic)
For the real monic quadratic , both roots lie strictly inside the unit disk exactly when , and . To justify this, a conjugate pair has squared modulus ; for real roots the first two inequalities rule out a single root beyond either endpoint, and the last rules out two roots beyond the same endpoint. Conversely the three inequalities follow directly from roots inside the disk.
Here and . The three quantities are , , and . They are all positive precisely for . At the roots are and . At , ; for example, , yields . If , the product of the roots is , so at least one root has modulus greater than one.
Worked examples
Worked example 4.3.4
(One actual step)
Take , , and . Then , , so and . The stored next slope is , not the predicted slope . The fully implicit trapezoidal step from the same instead solves , giving . Thus the two methods really do give different answers.
Worked example 4.3.5
(Stable oscillation and an unstable boundary)
For and , the roots are , both of modulus : all histories decay, possibly with sign oscillations. At , both roots equal , and some histories grow linearly. The trapezoidal rule instead has factor on this test equation, whose modulus is at most one for . The one-correction algorithm does not inherit that A-stability.
Worked example 4.3.6
(A refinement check with an exact starter)
Take , on , with the exact starting values and , so the theorem applies. Running the method to gives
The successive error ratios are about and , approaching the factor four expected for second order. Here lies well inside the stability interval, so propagation does not interfere. The ratios are evidence consistent with the theorem, not a proof of it.
Common misunderstandings
Common mistake
(Storing the predicted slope)
The predictor is an intermediate state, not the accepted new state. After the correction, must be evaluated and stored. In the one-step example the stored slope is , not the predicted slope ; reusing the latter defines a different method with a different stability analysis.
Common mistake
(Checking only the modulus of a boundary root)
At both roots have modulus one, so a test of modulus alone would accept this value. The root is double, the recurrence has the solution , and the value is unstable under the bounded-root convention.
Common mistake
(An arbitrary second starting value)
The convergence theorem assumes . Copying gives a starting error of order for a nonconstant solution, and the theorem can then no longer be invoked. A one-step starter such as explicit midpoint meets the hypothesis.
A small local defect establishes accuracy only when the accumulated errors are also controlled. Likewise, stability of the homogeneous test recurrence is not a general error bound for nonlinear equations, and the interval above applies to the specified one-correction formula with constant steps; additional corrections or changing step sizes give different recurrences.
Counterexample mode
The corrected method does not inherit A-stability
Since the corrector is the trapezoidal rule, one might expect the method to be stable for every with , as the trapezoidal rule is. Take with , so . Then and , and the characteristic equation has roots
From , the recurrence gives and , while the exact solution decays like . The trapezoidal factor at the same is . The repaired statement is the stability interval proved above: on the negative real axis the one-correction method is stable only for .
Quick check
Quick check
(Why can the corrector retain second order even though the predictor is not exact?)
Follow the predictor error through the endpoint slope.
Solution · Answer
Starting from exact past values, the predictor's endpoint error is . The endpoint slope error is multiplied by , contributing only to the corrected state. The underlying trapezoidal defect is already , so the exact-start defect stays .
Practice with solutions
Quick check
(For with , is the recurrence stable under the bounded-root convention?)
Compute .
Solution · Answer
Here , inside , so both roots lie strictly inside the unit disk and both modes decay.
Quick check
(Compute the first corrected step when , , and .)
Use and .
Solution · Answer
The predictor is , and the correction is . The slope to store is .
Quick check
(At , find the solution with , . What does it show?)
The characteristic polynomial is .
Solution · Answer
The recurrence has a double root , so ; the starting values give . An arbitrarily small starting perturbation grows without bound as increases.
Quick check
(A student uses an exact but chooses . Can the second-order convergence theorem be invoked?)
Compare the starting error with the hypothesis.
Solution · Answer
No. The second starting error is , not , and the theorem requires both starting errors to be . This fact alone does not determine the error of every later value; it only removes the guarantee.