From a differential equation to a sequence
An initial value problem specifies a rate of change and a starting state. Except in special cases, its solution cannot be expressed in elementary functions. A numerical method replaces the curve by values at a finite set of times. The resulting sequence must be judged on two different scales: whether it approaches the exact solution as the step is refined on a fixed time interval, and whether a chosen step controls growth over many iterations.
Consider
where and takes values in . Choose a positive integer , let , and write . We use one norm on throughout. A one-step method has the form with equal to the exact initial value. Each approximates ; it is not a continuous solution between grid points unless an interpolation is separately defined.
Euler's method and local defect
The tangent approximation gives explicit Euler:
The derivative is evaluated at the known state, so each step is explicit. To measure the quality of one step, start it from the exact solution.
Definition 8.1.1
(Unscaled local defect)
For a one-step method and the exact solution , the unscaled local defect of the step from to is
It is the error made in one step that starts at the exact value .
Some conventions divide this expression by ; here we do not. The defect differs from the global error , which accumulates defects and propagates earlier errors.
Suppose and . The vector integral remainder gives
Thus the defect is , or after division by .
Definition 8.1.2
(Order and consistency of a one-step method)
A method has order when its unscaled defect is uniformly over the relevant exact solution as . It is consistent when the defect divided by tends to zero.
In this convention Euler has order one. These statements concern a family of steps, not one fixed value of .
Worked example 8.1.3
(Euler on exponential growth)
For , , explicit Euler gives . With , the two values up to are and , so the error at is . With the four steps give and the error is about . Halving multiplies the error by about , which is already moving toward the factor expected from first order. For , as . This calculation shows convergence in this example; it is not yet a proof for general .
How local error becomes global error
Theorem 8.1.4
(Global error of explicit Euler)
Assume the exact solution is with on , and is continuous and globally Lipschitz in its state variable with constant , uniformly in . Starting from the exact , Euler satisfies
In particular, the global error is as with fixed.
Proof
(A recurrence for the global error)
Subtract the numerical step from the exact step. The local defect bound and Lipschitz condition imply
For , iterating the inequality and summing the geometric series gives . Since , this is the claimed bound. For , sum the defects directly and use .
Proof X-ray
Two jobs and one lost power of h
The recurrence separates two jobs. The bound on is used only for the fresh defect added at every step; the Lipschitz constant is used only for the factor that carries the previous error forward. The one non-obvious move is the final count: there are steps, each adding , so the sum is and one power of is lost. The geometric factor stays bounded because does not grow with . Without the Lipschitz hypothesis the amplification is uncontrolled; for example , has the exact solution , which does not exist on for .
Global Lipschitz continuity is a convenient sufficient hypothesis. Local versions can be proved when exact and numerical states remain in a suitable common region, but that containment must be established rather than assumed without comment.
Common mistake
(Local order is not global order)
Euler's local defect is , but its global error is only . Reading "defect of order two" as "second-order method" ignores the steps over which defects accumulate. Under the unscaled convention above, a method of order has defect and global error .
A two-stage Runge–Kutta method
Euler uses one slope. Averaging two slopes gives a method of higher order.
Definition 8.1.5
(Heun's method)
Heun's method, also called explicit trapezoidal Runge–Kutta, averages the starting slope with a slope predicted at the end of the step:
The endpoint in is an Euler prediction, not the unknown exact endpoint. The method is still explicit.
Worked example 8.1.6
(Heun's method on exponential growth)
For , we have and , so each step multiplies by . With this factor is , and two steps give , with error about at , against for Euler. Greater order is an asymptotic statement, not a guarantee that every coarse-step approximation is better in every problem.
Proposition 8.1.7
(Second-order accuracy of Heun's method)
In addition to the global Lipschitz condition, assume has continuous derivatives through order two on an open set containing, for sufficiently small , every line segment from to with , with derivatives through order two uniformly bounded on those segments. Then Heun's unscaled local defects are uniformly on and its global error on this fixed interval is .
Proof
(Matching the Taylor expansion through order two)
Starting from an exact value , write . Taylor expansion gives
The exact solution satisfies , so Taylor expansion of agrees with the Heun step through the term. Bounded derivatives make the remainder uniform, giving defect . For two starting states , the first slopes differ by at most ; the predicted endpoints differ by at most . Consequently the two Heun outputs differ by at most
The error recurrence has this amplification factor and an additive defect. For , the factor is at most for . Summing at most propagated defects gives on .
Definition 8.1.8
(Classical four-stage Runge–Kutta method)
The classical Runge–Kutta method uses
Under sufficient smoothness and appropriate uniform bounds it has global order four on a fixed interval. The full order proof requires matching additional Taylor terms and is not derived here. On , direct substitution gives the multiplier , which agrees with through degree four. That one test verifies the calculation for this equation, not the general order theorem.
Concept lensAnalytic
Runge–Kutta steps as quadrature rules
Since , each step approximates an integral. When depends on alone, Euler is the left rectangle rule , Heun is the trapezoid rule , and the classical method is Simpson's rule , because . When depends on , the stages are predicted values of the integrand.
Stability depends on the step
Consistency and convergence as do not say that a particular coarse step will behave well. Apply a method to the scalar complex test equation .
Definition 8.1.9
(Stability function and absolute stability region)
A one-step method applied to takes the form , where is its stability function. The absolute stability region is the set of for which . Strict inequality gives decay of this numerical mode, while equality gives bounded but generally nondecaying magnitude.
For Euler, , so its region is the closed disk . For Heun, . If is real, Euler has strict decay precisely when . At its multiplier is , producing bounded alternating values rather than decay. When , the exact solution decays, but outside the disk the numerical magnitude grows for nonzero initial data.
Worked example 8.1.10
(Euler on at three step sizes)
Here with , so . For the multiplier is : and every later value is , which is strict decay (the exact value is small but not zero). For , the endpoint , the multiplier is and , bounded but not decaying. For the multiplier is , so grows although the exact solution decays. Only gives strict decay.
Common mistake
(Boundary of the stability region is not decay)
Writing the stability condition as includes its boundary, but on the boundary the numerical solution need not decay. Euler on with has multiplier and produces forever, while the exact solution tends to zero. Use the strict inequality when decay is required.
Counterexample mode
A consistent method at an unsuitable step
Claim: a method that converges as produces a decaying numerical solution for a decaying problem, at whatever step is used. Take , and Euler with . Then , so and : the values reach at , that is at , where the exact solution is . Euler is still a consistent, first-order convergent method as on every fixed interval, so convergence and stability are separate properties. The repair is a step restriction: strict decay needs . A system with rapidly decaying modes can force an explicit method to take tiny steps for stability even when the desired solution changes slowly; this is one feature of stiffness, whose full theory needs a more precise problem-dependent definition.
Practice with solutions
Quick check
(Apply Euler and Heun with for two steps to , . Compare both with .)
Use the multipliers and .
Solution · Answer
Euler gives ; Heun gives . Both are below in this example.
Quick check
(For , determine the positive Euler step sizes giving strict numerical decay. What happens at the endpoint?)
Write the multiplier and require its modulus to be less than one.
Solution · Answer
Strict decay requires , or . At the multiplier is and the magnitude stays constant.
Quick check
(For Heun applied to with real , determine the positive steps giving strict decay.)
Put and study .
Solution · Answer
The multiplier is , which is positive for all real (its discriminant is ); it is less than one exactly when , that is, . Thus .
Quick check
(A method has unscaled defect bounded by and a one-step Lipschitz factor at most on a fixed interval. With exact initial data, what global order follows, and why?)
Write the recurrence for the global error and sum it.
Solution · Answer
The error satisfies with . At most terms, each amplified by at most for , give . Hence the global error is on the fixed interval.