Motivation
For a decaying differential equation, it is natural to expect a numerical solution to decay as well. A method can converge as the step size tends to zero and nevertheless grow at a step size used in a calculation. The distinction matters in stiff equations, where a rapidly decaying component can force an explicit method to take tiny steps even when the solution of interest changes slowly.
A two-step method also has a second numerical mode. This parasitic mode is not a second solution of the original first-order equation: it appears because the recurrence needs two starting values. We must track both roots, including multiplicities, before we can decide whether perturbations remain bounded.
Definitions
For , put , where is fixed. Applying a two-step method gives a homogeneous linear recurrence. A root of its characteristic polynomial represents a mode ; a repeated root also produces a mode .
Definition 4.2.1
(Absolute stability at )
A value is absolutely stable for a linear multistep method when every root of the method's characteristic polynomial at has modulus at most one and every root on the unit circle is simple.
This convention includes bounded boundary modes. Some texts use the strict interior, where all roots have modulus less than one, as the stability region; endpoints should therefore be stated explicitly.
The second-order Adams–Bashforth and backward differentiation formulas are, respectively,
On the test equation their characteristic equations, with the step-to-step factor, are
Definition 4.2.2
(A-stability)
A method is A-stable under this convention if every with is absolutely stable.
Zero stability is the root condition at ; it is necessary for convergence of a consistent multistep method, but says nothing by itself about an arbitrary nonzero .
Theorem/proposition
Theorem 4.2.3
(AB2 on the negative real axis)
For with , AB2 satisfies the stated root condition exactly when . Both roots lie strictly inside the unit disk for . At or the unit-modulus root is simple.
Theorem 4.2.4
(BDF2 is A-stable)
For every complex with , both BDF2 roots have modulus at most one, with any unit-modulus root simple. For in that closed half-plane both roots are strictly inside the unit disk; at the roots are and .
Proof
Proof
(Interval of AB2 from three sign conditions)
For AB2, the polynomial is , with and . For a real monic quadratic, both roots lie strictly inside the unit disk exactly when
To see this, the first two expressions are and . If the roots are a complex conjugate pair, their product is their common squared modulus, so gives the claim. If they are real, positivity of and excludes one root inside and one outside either endpoint; if both were outside on the same side their product would exceed one. Conversely, two roots strictly inside plainly give these three inequalities. Here the conditions reduce to , , and , establishing the open interval. At the roots are ; at they are . If , then , while as , so a real root is less than . The endpoint is bounded but does not decay.
Proof
(Reciprocal roots for BDF2)
For BDF2, suppose a root has . It is nonzero, and satisfies . Dividing the characteristic equation by gives
Write . Since ,
Equality throughout requires , hence , and . Thus no root can have modulus at least one when and . At the roots are and , both simple. Also in the closed left half-plane, so the equation remains quadratic there. This proves the assertion without assuming the two roots are real.
For example, at the BDF2 roots coincide at ; the additional mode still decays. Only roots on the unit circle must be simple.
Proof X-ray
Why inverting the root makes BDF2 transparent
The characteristic equation is linear in , so instead of solving for the proof solves for . Substituting turns the region to be excluded, , into the closed unit disk , and expresses as the explicit quadratic . The single inequality is where is used; it bounds below by , a quantity that cannot be negative. The hypothesis enters only at the end, where it forces equality and hence , . That leftover case is closed separately by computing the roots and and checking simplicity. The same substitution applied to AB2 gives ; at this equals , a point of the left half-plane, which is exactly why AB2 cannot be A-stable.
Worked examples
Worked example 4.2.5
(A stable and an unstable AB2 step)
For and , and the AB2 roots are . Their moduli are below one, so every starting pair produces a decaying sequence. With , the roots become ; the root has modulus greater than one. Roundoff or starting error can excite it even if a specially chosen starting pair happens to suppress it.
Worked example 4.2.6
(BDF2 for the same stiff mode)
At , BDF2 gives , with roots . Each has modulus . Both numerical modes decay, although the roots are complex and the computed solution can oscillate. Stability guarantees bounded propagation, not exact reproduction of the positive factor .
Worked example 4.2.7
(A parasitic starting error)
At , exact solutions are constant. Take BDF2 starts and . Solving the recurrence yields . The transient mode decays, but a shifted constant mode remains: zero stability does not erase a start error. A second-order computation needs starting values of suitable accuracy.
Common misunderstandings
Common mistake
(Reading A-stability as an accuracy guarantee)
A-stability is a test-equation property, not an unconditional accuracy guarantee or an automatic nonlinear error bound. A large step may damp a component while resolving its transient badly, as the oscillating BDF2 modes at show.
Common mistake
(Treating a finite stability interval as non-convergence)
AB2's finite stability interval does not contradict convergence on a fixed time interval: as with fixed , approaches zero and eventually enters the interval .
A further error is to examine only the root that approximates . The parasitic root has no counterpart in the differential equation, yet it can control long-run growth.
Counterexample mode
Checking only the principal root
Claim: if the AB2 root that continues from has modulus less than one, the step is stable. Along the discriminant is positive, so the two roots stay real and distinct, and the larger one moves continuously from . At the characteristic equation is , whose roots are and . The principal root decays, but the parasitic root has modulus about , so the starts , give . The correct test is the root condition for every root of the characteristic polynomial.
Quick check
Quick check
(Why is stable under the convention above for AB2, but unsuitable if one needs strict numerical decay?)
Compute both roots at .
Solution · Answer
At the AB2 roots are and . Both are simple and have modulus at most one, so the root condition holds. The mode , however, alternates forever with unchanged magnitude, so a perturbation in that mode never decays.
Exercises
Quick check
(Determine whether AB2 is stable for with and with . Give the reason at the endpoint.)
Solution · Answer
The products are and . The first satisfies the root condition: its roots and are simple, and the root on the unit circle is bounded but does not decay. The second is unstable because ; indeed , so a real root lies below .
Quick check
(At , calculate both BDF2 roots and their common modulus.)
Solution · Answer
The equation is , so and .
Quick check
(A student tests BDF2 at and concludes it is A-stable. Explain the gap.)
Solution · Answer
At the roots establish only zero stability. A-stability requires the root condition for every complex in the entire closed left half-plane; the reciprocal-root argument above supplies that missing proof.
Quick check
(For AB2 at , give initial values that isolate the growing mode, and explain why this does not disprove convergence as for fixed .)
Solution · Answer
Let and take , . Then satisfies the recurrence and grows. For a fixed , reducing changes ; eventually , and finite-time convergence is a separate assertion involving consistent starting values.