Why look beyond an explicit step?
A stable differential equation can decay rapidly in one direction while changing much more slowly in another. An explicit method may then need a small step merely to keep the fast numerical component from growing. The following two implicit methods provide a different stability tradeoff. Their stability on a scalar test equation is useful evidence, not a complete definition or diagnosis of stiffness for every nonlinear system.
Consider at times with fixed .
Definition 8.3.1
(Backward Euler and the implicit trapezoidal rule)
Backward Euler and the implicit trapezoidal rule define by
Unlike an explicit update, these are equations to solve for the new value.
Common mistake
(The implicit trapezoidal rule is not Heun's method)
Heun's method has the same averaged form, but it evaluates the end slope at a predicted value , so it is explicit. The trapezoidal rule above evaluates the end slope at the unknown itself. The two share the order of accuracy, but their stability functions differ: for Heun gives , which is unbounded in the left half-plane, whereas the trapezoidal rule gives a bounded rational function.
Scalar stability functions
For , put . Whenever their denominators are nonzero, the two methods give with
Definition 8.3.2
(A-stability and L-stability)
A method is A-stable here if for every in the closed left half-plane , with defined there. It is L-stable if it is A-stable and as within the closed left half-plane.
For a strictly decaying mode, , these two rules even give strict inequality. Boundary equality may preserve an undamped oscillation; it is not numerical decay.
Theorem 8.3.3
(A-stability of backward Euler and the trapezoidal rule)
Both and are defined, with magnitude at most one, at every with . Hence both methods are A-stable.
Proof
(Comparing squared moduli in the left half-plane)
For backward Euler, in the closed left half-plane. Its denominator is therefore nonzero there and . For the trapezoidal rule,
Its denominator is nonzero there, since ; its amplification factor therefore has magnitude at most one.
Proof X-ray
Where the sign of the real part enters
The hypothesis is used twice, for two different purposes. It makes , which keeps the denominator away from zero so that is defined; and, for the trapezoidal rule, it makes the difference of squared moduli nonnegative, which is the inequality . The non-obvious move is to compare squared moduli of numerator and denominator instead of computing in real and imaginary parts. On the imaginary axis the difference is exactly zero, so there: the proof gives no strict decay at the boundary, and it does not hold at all for , where the same difference is negative.
This proof applies to the test equation and exact solution of the implicit step; it is not by itself a nonlinear stability theorem.
Backward Euler is L-stable since ; the trapezoidal factor instead tends to along the negative real axis. Thus trapezoidal steps do not strongly damp arbitrarily fast stable scalar modes, even though their magnitude never exceeds one in the left half-plane.
Counterexample mode
A-stability does not make fast modes die
Claim: if a method is A-stable, a very fast decaying mode is strongly damped after a few steps of any size. Take with . The trapezoidal rule gives . After ten steps the magnitude is , so the numerical value has barely decayed and still alternates in sign, although the exact solution has shrunk by the factor . Backward Euler gives at the same . The repair is the stronger hypothesis L-stability, which requires as ; backward Euler has it, the trapezoidal rule does not.
A fast-decay comparison
Worked example 8.3.4
(One step at )
Take , and , so . One step gives:
| Method | Amplification | Numerical |
|---|---|---|
| Explicit Euler | ||
| Backward Euler | ||
| Implicit trapezoidal |
The exact value at is , approximately . Explicit Euler grows in magnitude on repeated steps; backward Euler and the implicit trapezoidal rule both decay for this . After two steps the three values are , and . The trapezoidal values alternate in sign, even though the exact solution stays positive. A-stability therefore guarantees neither positivity nor small error at a large step. For sufficiently large negative real , backward Euler's damping can also erase a fast transient that matters to the problem.
Worked example 8.3.5
(A pure oscillation on the boundary)
For the exact solution has modulus one for all . Take , so . Backward Euler gives , of modulus : the numerical oscillation is damped by that factor every step. The trapezoidal rule gives , of modulus exactly one, so the amplitude is preserved. Its phase advance per step is , while the exact advance is , so a phase error remains.
Accuracy and solving the implicit equation
If the exact solution is sufficiently smooth, substituting it into a backward Euler step leaves a one-step residual ; trapezoidal quadrature gives a residual . Under suitable regularity, well-posed implicit solves, and stability assumptions on a fixed finite time interval, these lead to global orders one and two respectively. These are small-step accuracy statements; neither follows merely from A-stability. An algebraic solver also introduces error: a loose stopping tolerance can dominate a nominally higher-order discretization.
Worked example 8.3.6
(Backward Euler as an equation for the new value)
Implicit does not automatically mean a unique real update. Apply backward Euler to at with . The new value must satisfy , that is , giving . Both are real: and . The root near the old value is the one consistent with the small-step continuation. Newton's method or fixed-point iteration needs a specified initial guess, stopping criterion and failure policy; convergence is not guaranteed by writing down an implicit formula. Conditions such as a contraction on a chosen region for small enough can guarantee a unique root there, not necessarily a unique root over all real numbers.
Common mistake
(A-stable does not mean accurate)
A-stability only bounds the amplification factor for . At a large step the numerical value can have the wrong sign, as the trapezoidal value against shows, and a method can be A-stable with order one only. Accuracy is a small-step property of the defect; stability is a statement about the amplification of the whole step. Neither implies the other.
Exercises
Quick check
(For , compute the backward Euler and trapezoidal amplification factors and compare their signs with the exact factor .)
Evaluate and .
Solution · Answer
and . The exact factor is positive, so backward Euler has the correct sign and the trapezoidal factor has the wrong one.
Quick check
(Prove that for , real , trapezoidal amplification has modulus one. Explain why this does not establish exact phase.)
The numerator and denominator are complex conjugates.
Solution · Answer
, and the numerator is the complex conjugate of the denominator, so both have the same modulus and . Modulus one says only that the amplitude is preserved. The exact multiplier has argument whereas has argument , so the phase differs for .
Quick check
(On , , , determine whether the backward Euler equation has a real solution. What does this show about treating implicit steps as automatically solvable?)
Write the equation for the new value .
Solution · Answer
The equation is , that is , with discriminant . It has no real solution. An implicit step is an equation whose solvability, uniqueness and root selection must be checked, not assumed.