From a differential equation to an update
Suppose a model gives a rate of change rather than the future state directly. An initial-value problem specifies both:
Assume for the moment that a solution exists on the interval. We want approximations on the uniform grid , with . Integrating the equation exactly over one step gives
Replacing the integral by times its left endpoint value gives the first method.
Definition 1.9.1
(Explicit Euler method)
For a step size and the grid , the explicit Euler method starts from and sets
The current state determines the slope and hence the next state. Geometrically, one follows the tangent line for one time step. This is a numerical rule, not an assertion that the true trajectory is a straight line. A smaller step generally reduces discretization error, but more steps cost more work and floating-point roundoff can also accumulate.
Local defect and global error
The one-step defect measures what happens when the exact solution, rather than an already approximate value, is inserted into the numerical update.
Definition 1.9.2
(One-step defect)
Let be an exact solution on . The one-step defect of explicit Euler at step is
If is twice continuously differentiable and on , the integral form of Taylor's theorem yields . This is an absolute one-step defect of order ; some conventions divide it by and call the result the local truncation error. Stating the convention prevents an apparent disagreement about the order.
Theorem 1.9.3
(First-order global convergence of explicit Euler)
Let be a twice continuously differentiable solution on , with . Suppose is Lipschitz in its second argument with constant throughout a domain containing the exact and numerical values. Starting from the exact initial value, explicit Euler satisfies
Thus the error at a fixed final time is as the grid is refined, provided the stated bounds hold uniformly.
Proof
(Error recursion and a geometric sum)
Put . Subtracting the numerical update from the exact step gives
The triangle inequality, Lipschitz condition and defect bound imply
Since , iteration bounds the error by . When , summing the geometric series and using gives the displayed bound. When , the sum is , giving . The proof reveals why per step becomes only after roughly steps.
Proof X-ray
Where the convergence estimate gets its assumptions
The proof has one exact decomposition and one recursion. The decomposition splits into the propagated old error and the fresh defect ; nothing has been estimated yet. The Lipschitz hypothesis is used only on the propagated part, and it applies because both values of are taken at the same time . That is why the theorem needs a single constant on a region containing both the exact and the numerical values. The bound is used only on ; neither hypothesis can replace the other.
The key move is to stop estimating step by step and unroll the linear recursion from : the defect made at step is amplified by at most before it reaches step . The argument closes with , which turns amplification factors into the step-independent factor ; this is where a fixed final time is needed. If is only locally Lipschitz, one must first show that the numerical iterates stay in a common bounded region for small before one constant is available.
Worked example 1.9.4
(Refining an exponential calculation)
For , , the exact solution is . Euler gives . At and , two steps give , below . At , four steps give . The exact value is approximately . The second answer is closer, but two computations alone do not prove an error order; the theorem supplies a bound under its hypotheses.
The errors are about and , a ratio of about when is halved; the expansion shows that the error is about for small , so the ratio tends to , consistent with first order.
Common mistake
(One-step order is not global order)
A defect of size in each step does not give a global error of size . About defects accumulate, and the proof shows the total is ; for the global error is exactly , as the last practice item shows. When quoting an order, say whether it refers to the one-step defect, to the defect divided by , or to the global error.
Stability asks a different question
A convergence theorem holds as for a fixed problem and fixed time interval. In practical computations is finite. To see how errors propagate, examine the scalar test equation with complex and . Its exact solution decays. Explicit Euler instead multiplies the state and any perturbation at each step by
It reproduces decay over arbitrarily many steps exactly when . For real negative , this means . At perturbations need not decay; outside the disk they grow. This is an absolute-stability condition for the test equation, not an error bound for every nonlinear problem.
Counterexample mode
A decaying solution can produce a growing Euler sequence
Take , , and . The exact values are positive and decrease, but Euler gives , so and . The iterates alternate and grow. The method remains convergent as for this same problem; the chosen step simply lies outside its stability interval .
Implicit Euler and stiff decay
Using the right endpoint in the integral instead leads to a second method.
Definition 1.9.5
(Implicit Euler method)
For a step size , the implicit (backward) Euler method starts from and determines each as a solution of
The unknown now occurs on both sides. Each step requires solving an equation, which may be nonlinear and may not have a unique solution without further conditions. For example, if is globally Lipschitz with constant and , the right-hand side is a contraction in the unknown, giving a unique step by the contraction theorem. This is a sufficient condition, not a necessary one.
For the test equation the implicit equation is linear and can be solved in closed form.
Proposition 1.9.6
(Implicit Euler damps every decaying test mode)
Let with , and let . For , every implicit Euler step is uniquely solvable, , and whenever .
Proof
(The denominator has real part greater than one)
With the step reads . Since , the number is nonzero, so the step has the unique solution . Moreover , hence , and when .
The hypothesis places no upper bound on . For and , the factor is rather than the explicit factor . Stability does not guarantee accuracy: after one step is far from the exact . A large step can misrepresent a fast transient even while remaining stable.
Concept lensGeometric
Two stability regions in the complex plane
In the variable , explicit Euler damps the test mode exactly on the open disk of radius one centered at . It meets the negative real axis in , which is the condition . Implicit Euler damps it exactly where , outside the closed disk of radius one centered at . That exterior contains the whole open left half-plane, which is the proposition, and also every with , where a growing mode is wrongly damped.
A system can contain fast and slow decaying modes simultaneously. For instance, and have very different decay time scales. Explicit Euler must use to damp the fast mode even if the slow mode is the quantity of interest over a much longer interval. This is a simple manifestation of stiffness: a stability restriction can force much smaller steps than the desired resolution of the slow behavior. Implicit methods trade a less restrictive test-equation stability region for linear or nonlinear solves; neither method removes the need for accuracy checks.
Worked example 1.9.7
(An implicit step that must select a root)
Take , , whose exact solution is , and . Implicit Euler requires , that is
Both roots satisfy the step equation, but only continues the trajectory; has nothing to do with the solution. A practical solver starts an iteration at the previous value: Newton's method for from gives , converging to the correct root. Compared with , the implicit error is about , while explicit Euler gives , with error about .
Worked example 1.9.8
(One step size for a stiff pair)
For , and , the explicit factors per step (left) and the implicit factors (right) are
The exact slow factor is , so both methods resolve the slow mode reasonably at this step. Explicit Euler, however, multiplies any fast component by per step: after ten steps a fast perturbation of size has become about , since . Implicit Euler shrinks it by per step, although the exact fast factor is far smaller still. The boundary case shows that the explicit threshold is sharp: at the fast factor is , so a fast perturbation persists forever with alternating sign, neither growing nor decaying.
Common mistake
(Unconditional stability is not accuracy)
The proposition says that implicit Euler never amplifies a decaying test mode; it says nothing about how accurately it follows that mode. With and the factor is stable, while the exact factor is . The same property can also mislead for growing modes: if and , then , so implicit Euler damps a solution that in fact grows.
Practice
Quick check
(For , , use explicit Euler with to compute at . Compare with the exact answer.)
Evaluate the slope at the left endpoint of each step.
Solution · Answer
The slopes at and are and . Hence and . The exact solution is , so the error at is .
Quick check
(For , , compute two implicit Euler steps with . Compare with the explicit result above.)
Each step is the linear equation .
Solution · Answer
Each step solves , and , so and . The explicit answer lies below , whereas lies above it. This growing mode is not damped: implicit Euler does not suppress a solution that genuinely grows when .
Quick check
(For , which real positive step sizes make explicit Euler damp perturbations? What happens at the nonzero boundary?)
Write down the amplification factor .
Solution · Answer
The factor is . The inequality is equivalent to . At the factor is , so perturbations alternate in sign without decaying.
Quick check
(Let with an exact initial value. Show directly that the accumulated explicit Euler error at is in magnitude.)
Sum the left-endpoint slopes .
Solution · Answer
The exact solution is , while Euler sums left endpoints: . Subtracting from gives . Here and , so the theorem's bound is attained exactly.