A linear multistep method reuses old solution values and derivatives. Its coefficients may look plausible, but how can we tell the accuracy order without solving a particular differential equation? Insert an exact smooth solution, expand at one grid point, and require the coefficients of successive powers of the step to vanish. This gives a finite set of algebraic tests. It does not, by itself, prove that errors remain controlled over many steps.
Defect and coefficient moments
Definition 8.4.1
(Linear multistep method and unscaled defect)
For a fixed integer , a constant-step linear multistep method is
with real coefficients and . If , the new value is generally implicit. For an exact smooth solution , its unscaled defect is
Common mistake
(Scaled and unscaled defects differ by one order)
Some conventions divide (2) by . A defect in the unscaled convention is in the scaled one, and both describe a method of order . Quoting a defect exponent without saying which convention is used leads to an off-by-one error in the order.
Definition 8.4.2
(Coefficient moments)
For the moments of the method are
At , the factor is one even for . The moments depend only on the method, not on the solution.
Theorem 8.4.3
(Algebraic conditions for order)
For an integer , the following are equivalent:
- .
- For every sufficiently smooth exact solution, its unscaled defect is locally as .
More precisely, the implication from 1 to 2 holds uniformly for grid points on a fixed compact interval if there. If these moments vanish, then
for , uniformly on a fixed compact interval where all required points lie. If , testing a suitable polynomial shows that order does not hold in general. Consistency means order at least one: .
Proof
(Taylor expansion and a test for necessity)
Expand through degree and through degree . Bounded derivatives through order on the compact interval give a uniform remainder after the factor in (2). Collecting terms yields
Thus vanishing moments prove the defect estimate. Expanding one term further with the Peano remainder gives (4); uniform continuity of on the compact interval makes the little-o remainder uniform.
Conversely, suppose every sufficiently smooth exact solution has defect . For each , take the polynomial solution of the valid differential equation for , and a constant solution for . At its defect is exactly for , or for . Since , the asserted bound forces . The same test with degree proves the final assertion when .
Proof X-ray
Why polynomial test solutions isolate one moment
The two directions use different tools. Sufficiency needs smoothness: on a compact interval bounds the Taylor remainder uniformly. Necessity needs only exact solutions of valid equations, and the polynomials are the right ones because each equals its own Taylor polynomial at , so no remainder appears. At the defect is therefore exactly , with no remainder, and is isolated from all other moments. A defect is then incompatible with for any , since is not . The same test at degree shows why order fails exactly when . The moments say nothing about the roots of , which the next section shows to be a separate hypothesis.
The first two conditions can also be written and for and . They do not require to be nonzero; zero-stability is a separate root condition.
Examples: second-order methods
Worked example 8.4.4
(Two-step Adams–Bashforth)
Index the method as
so and . Then , and , while . Its unscaled defect therefore starts with when that derivative is continuous. In particular the method has order exactly two.
Worked example 8.4.5
(Implicit trapezoidal method)
For , use , and . Then , and , but . Its unscaled defect begins with . Equal order does not mean equal defect constant, implicitness, or absolute-stability region.
Worked example 8.4.6
(Two-step backward differentiation formula)
The formula is normalized by dividing by :
so and . Here , so the method is implicit, and . We get , , and . The unscaled defect begins with , so the order is exactly two. As a check against the root condition, and .
Common mistake
(Dropping the factor q in the moments)
In the second sum carries the factor from differentiating . For the trapezoidal method, and . Forgetting the factor gives and suggests only order one; with the factor, , the correct conclusion being order two.
Why consistency alone is not enough
Counterexample mode
A consistent method with a repeated unit root
Consider
Here and . Direct calculation gives , so it is consistent and even has order at least two. Its characteristic polynomial is , however, and the repeated unit root violates zero-stability.
Indeed, for with exact solution , choose , . Both starting errors tend to zero, yet the recurrence gives . At indices with , the error tends to , not zero. A small defect does not prevent an unstable recurrence from amplifying starting errors.
Three ingredients therefore answer three different questions. The moments test the defect of the exact solution. The root condition tests the homogeneous recurrence when . The starting values determine the initial error. The earlier zero-stability theorem bounds their combined effect: under its Lipschitz and small-step hypotheses, a zero-stable method of order with starting errors has global error on a fixed interval. For an implicit method, the new value must also be well-defined; a global Lipschitz condition and sufficiently small make its fixed-point equation a contraction. The order calculation alone supplies none of these extra facts.
Practice with solutions
Quick check
(For , calculate and determine its order.)
Read off and first.
Solution · Answer
, , so and . The order is exactly two, with leading defect .
Quick check
(For forward Euler , find and its leading unscaled defect term.)
Here and .
Solution · Answer
, so the leading term is .
Quick check
(If a method has , must be nonzero? Use (6).)
Compute for (6).
Solution · Answer
No. For (6), and , even though both consistency moments vanish.
Quick check
(A zero-stable order-two method starts with errors , rather than . What order does the previously proved error bound guarantee?)
Add the propagated starting error to the accumulated defects.
Solution · Answer
The zero-stability estimate gives on a fixed interval; it does not guarantee the method's full second order with this starter.