Reusing past derivative evaluations
A one-step method advances from using information generated at the current step. A multistep method also retains earlier values. For the differential equation :
Definition 8.2.1
(Two-step Adams–Bashforth method)
The two-step Adams–Bashforth method (AB2) is
It needs two starting values . Once they are available, one new evaluation of per step suffices if the previous value is stored. The negative coefficient does not mean time is running backward: it comes from extrapolating the derivative over the next interval.
Derivation and local defect
Let along a sufficiently smooth exact solution. Interpolate linearly at and :
Integrating from to gives , the AB2 increment.
Definition 8.2.2
(Unscaled local defect of AB2)
The unscaled local defect is the error of one step started on the exact solution:
Proposition 8.2.3
(Size of the AB2 local defect)
If has a bounded second derivative, the defect is uniformly on a fixed interval. If locally, then .
Indeed, expand and to second order and compare their integrals: the terms in and cancel, and the second-order terms leave .
Common mistake
(Local truncation error and unscaled defect differ by a power of h)
Some books divide the defect by and call the result the local truncation error; under that convention its order is . The defect used here is . Quoting either exponent without stating the convention causes a common off-by-one disagreement about the order.
Worked example 8.2.4
(One AB2 step with an exact starter)
Take , , and the exact starting pair , . Then . The exact value is , so the error is about . Since the starting pair is exact, this error is the defect, and the leading term predicts , within about one percent.
From a defect to a global error
Local accuracy alone does not establish that repeated steps stay accurate. The following elementary argument proves convergence for this particular method.
Theorem 8.2.5
(Second-order convergence of AB2)
Suppose is Lipschitz in its second variable with a global constant for all state values and , and the exact solution has a uniform defect bound up to a fixed final time . If , then the AB2 values satisfy as .
Proof
(A two-term error recurrence)
Put and for . Subtract the exact recurrence from the numerical one. Lipschitz continuity and the defect bound give
The same upper bound also dominates , so . Iteration for at most steps bounds the product by and the accumulated defects by a constant times . Since , the conclusion follows. The same proof works for vector-valued systems with a norm and a Lipschitz constant.
Proof X-ray
The starting value is part of the algorithm
The recurrence involves two earlier errors, so the proof tracks the maximum of the last two; this is what turns the two-term recurrence into a one-term inequality, with the Lipschitz constants and adding up to . An defect occurs at every step, and there are steps on a fixed interval, which costs one power of . The product stays bounded, so stability of the error recurrence prevents amplification from destroying the resulting bound. The starting error enters only through , and it is multiplied by the same bounded constant: a starting error that is only is not covered by this second-order theorem, even when every later AB2 step is computed exactly.
Worked example 8.2.6
(An Euler starter on exponential growth)
For , , , take the starter from one explicit Euler step; its error is , which is . Then , and the error against is about . This is more than ten times the error obtained from the exact starter in the previous example: the second-order starter is good enough for second-order convergence as , but its error dominates the constant at a fixed step.
Absolute stability on a test equation
Apply the method to , and put . A trial gives
Definition 8.2.7
(Absolute stability of AB2 at a given )
For a fixed , absolute stability requires the numerical solutions to decay for arbitrary starting pairs; equivalently both roots of the polynomial above have modulus strictly less than one.
A repeated root also produces a factor , but still decays when its modulus is below one. At the roots are and , so the numerical solution remains bounded but does not decay. This boundary case is called zero-stable behavior, not strict absolute stability.
Proposition 8.2.8
(Negative real stability interval of AB2)
For real , both roots of the AB2 stability polynomial lie in if and only if . At one root is , and for one root has modulus greater than one.
Proof
(Signs of the polynomial at one and minus one)
For real negative , write with and . The roots are real because the discriminant is . Their product is , so one root is negative and one positive. Evaluating the characteristic polynomial at gives , while at it gives . Since with , the positive root is below one exactly when , which holds for every . Since with , the negative root exceeds exactly when , that is . It follows that both roots lie in exactly when .
Common mistake
(The stability interval is open)
Writing the condition as includes the endpoint , where the roots are and . The root gives a component that never decays, so strict absolute stability fails there even though the numerical solution stays bounded.
Worked example 8.2.9
(Three step sizes for )
Here and strict stability requires . For , and the polynomial has roots and , both inside the unit interval. For , is the boundary: the roots are and , and with , the solution is , whose alternating part never decays. For , and the roots are , about and ; one has modulus greater than one. The exact solution decays, but a generic numerical starting pair excites the growing mode. Choosing a small local defect does not remove this step-size restriction for the long-time test problem.
Counterexample mode
An accurate formula can still amplify perturbations
Claim: since AB2 has a small local defect, it gives a decaying solution for the decaying equation at any moderate step. Take , so and the recurrence is . Start from the exact pair , . Then , , and , the ratios approaching , whereas the exact value at is . Even a tiny component in the unstable root's direction grows geometrically with the number of steps. This is a statement about fixed and long-time behavior for a test equation. It does not contradict the finite-interval convergence theorem, where and for fixed . The repair is the step restriction .
Practice and solutions
Quick check
(For , starting from , , use AB2 once to find .)
Evaluate at both starting values.
Solution · Answer
The increment is , so .
Quick check
(Find the two characteristic roots at and decide whether every numerical solution decays.)
Substitute into .
Solution · Answer
The polynomial is , with roots , about and . Both have modulus less than one, so every numerical solution decays.
Quick check
(If the exact starting value has but , does the convergence theorem above guarantee second-order accuracy?)
Compare the starting error with the hypothesis .
Solution · Answer
No: the assumed starting error is , whereas this error is . The theorem supplies no second-order guarantee.
Quick check
(For , give the strict negative-real stability restriction on .)
Write and use the interval .
Solution · Answer
Here , so means .