Motivation
A differential equation describes a rate of change at every instant. A computer stores only finitely many values. Numerical integration bridges these descriptions by replacing the continuous evolution with a recurrence on a time grid. The recurrence is useful only if we understand how its approximations relate to the original solution. A smooth-looking list of numbers provides no error certificate by itself.
Consider an initial-value problem
The unknown may be a vector in . We assume that a classical solution exists on this entire interval; existence must be checked separately when the equation can blow up. At a grid point, the differential equation supplies the tangent vector. Following that tangent for a short time suggests forward Euler's method. The important question is what happens after many such steps, because each new tangent is evaluated at an approximate state.
Two mechanisms are involved. Even when a step starts at the exact solution, a tangent line misses the curved solution. Also, a perturbed starting value can change later values. The first mechanism is measured by consistency; the second is controlled by stability with respect to perturbations. A convergence proof combines them. Keeping their roles distinct is useful far beyond Euler's method: the same reasoning underlies Runge–Kutta methods and time discretizations of PDEs.
Definitions
Definition 1.1.1
(Forward Euler, global error and one-step defect)
Choose an integer , set , and write . Forward Euler computes
Capital denotes a computed approximation; lower-case denotes an exact value. Fix one norm on . The global error is . Its size at a fixed final time is a different quantity from the error made in just one step.
The unscaled one-step defect of the exact solution is
The scaled defect is . Some books call either quantity the local truncation error. Always state the convention: Euler has and under the smoothness assumptions below. The powers differ because the quantities differ, not because the method has two competing orders.
Definition 1.1.2
(Order of convergence and the Lipschitz condition)
A bound means that there are constants , independent of the mesh size, such that whenever . For an error sequence, a uniform bound means that the same works for every grid point up to the fixed time . Convergence of order at least one means
for sufficiently small . The constant may depend on the equation and on . It must not secretly grow with .
We use a Lipschitz condition in the state variable:
In the theorem it holds for every and all , with one constant . This global assumption avoids having to prove that the numerical states stay in a smaller region. Local Lipschitz bounds can also be used, but then confinement of both exact and numerical trajectories needs justification.
Concept lensAnalytic
Euler as a left-endpoint quadrature rule
Integrating the equation over one step gives the exact identity
Forward Euler replaces the integral by the left rectangle , with the integrand frozen at the computed state. The defect is exactly the rectangle rule's error for the integrand ; the Lipschitz constant controls the further change caused by using in place of . When does not depend on , Euler is precisely a left Riemann sum.
Theorem/proposition
Proposition 1.1.3
(The exact one-step defect)
Suppose and throughout the interval. Then forward Euler's exact-solution defect satisfies at every step.
Theorem 1.1.4
(A finite-time global error bound)
Suppose is continuous, is globally Lipschitz in its state variable with constant , and has a solution with . With exact initial data, forward Euler satisfies, for ,
For the corresponding bound is . Thus the method converges uniformly to the solution on the fixed interval as .
Proof sketch or proof idea
Proof
(Integral Taylor formula for the defect)
The fundamental theorem of calculus, applied twice, gives an integral Taylor formula:
Substitute . Taking norms in the remaining integral proves the defect estimate because the integral of is . The integral proof works for vectors without selecting a single intermediate Taylor point for all components.
Proof
(A scalar recurrence for the global error)
Subtract the numerical recurrence from the exact identity. Since
the Lipschitz and defect bounds yield
Starting with and iterating this scalar inequality gives a geometric sum. If , that sum is . If , it is . Finally use for . Every step in the argument uses an upper bound; it does not assert that the actual errors align in the worst possible direction.
Proof X-ray
Where the two hypotheses enter
Smoothness is used once, in the defect proposition, and only along the exact solution: the integral Taylor formula turns into . The Lipschitz condition is also used once, to compare the slopes and at the exact and computed states. That comparison produces the amplification factor . Because the condition is global, no separate argument is needed to keep in a region where a local constant is valid.
The key move is to replace the vector recurrence by the scalar inequality and solve it exactly as a geometric sum. About defects of size are added, and each is amplified by at most , a factor independent of . Multiplying the local defect by the number of steps without this factor ignores amplification and is not a proof. Without the Lipschitz bound the factor cannot even be formed, and the counterexample under Common mistakes shows that the conclusion can then fail.
If the initial value has error , the same calculation adds to the bound. If step also suffers an additive perturbation , its contribution is bounded by
This formula covers solver inaccuracies and roundoff at the level of a specified perturbation model. A fixed per-step absolute perturbation need not shrink when does, so reducing the step indefinitely does not guarantee indefinitely improving floating-point accuracy.
Worked examples
Worked example 1.1.5
(Four tangent steps for exponential decay)
For , , Euler gives . With , the values are , , , , and . At , the exact value is , so the signed error is approximately .
Here and , so the theorem gives the valid, looser bound . A rigorous upper bound is not a prediction of the observed error. For , , and the error is approximately . The ratio is about , approaching the factor two expected from first-order convergence.
Worked example 1.1.6
(An exact defect with accumulated error)
For , , the exact solution is . The exact defect is . Summing the numerical increments gives
In this problem is independent of the state, so , and . The global estimate is attained exactly. There is no amplification of previous errors, but the new defects still accumulate. The example also distinguishes a local defect, , from the final error, .
Worked example 1.1.7
(Choosing a step from an explicit tolerance)
For , , on , require a proven absolute error at every grid point of at most . The estimate gives , or . Taking gives and a bound approximately .
This is a sufficient choice, not the smallest possible number of steps. The proof uses a worst-case Lipschitz estimate which does not exploit decay. An empirical run with fewer steps might satisfy the tolerance, but that observation has a different logical status from this guaranteed bound.
A refinement experiment and its limits
When the exact answer is unavailable, compare results at a common time using , , and . If with , then tends to two. The derivation uses an asymptotic expansion; a single ratio near two is evidence of the expected regime, not proof that every component has reached it. If the leading coefficient vanishes, a different ratio can occur even for the same method.
For this expansion, cancels the leading error. This extrapolation can improve an approximation when the expansion is valid, but an order bound alone does not imply such an expansion. Distinguish what a theorem establishes from what an experimental error model assumes.
Obtaining the constants rather than assuming them
An error estimate becomes useful when its constants can be bounded from the equation. If has continuous first derivatives, the chain rule along a solution gives . Thus a region containing the exact trajectory can supply a bound for by bounding these three quantities there. This does not require knowing the exact solution formula. A differential inequality, a conserved quantity or an a priori bound may identify such a region. If no such region is available, quoting a numerical value for without justification leaves a gap in the guarantee.
For a vector equation, a bound on the operator norm of implies the Lipschitz inequality on a convex region: integrate the derivative along the straight segment joining two states. A bound on individual entries of the Jacobian may be converted into an operator norm bound, but the conversion depends on the norm and dimension. Mixing a maximum norm estimate for one step with a Euclidean norm estimate for another can silently change the constant. Fixing the norm at the start makes the argument traceable.
The error recurrence can also be checked carefully by induction. After the first step only the new defect is present; after the second, the error contains the first defect amplified once together with the newly added second defect. The further the iteration proceeds, the more propagation steps the earlier defects have undergone. Taking norms and applying the triangle inequality discards information about cancellation between signs, which is exactly why the upper bound is sometimes far larger than the observed error. To exploit a decaying direction or cancellation of errors, one must exhibit additional structure; a general proof cannot simply assume that it is present.
When reporting results in practice, record both the number of steps and the final time. Before comparing final values from two grids, confirm that they refer to the same time; if an integer number of steps truncates the run at an earlier time, the difference also mixes in a discrepancy in time position. If the last step is deliberately shortened to reach a prescribed final time, use the nonuniform-step analysis below rather than treating the last interval as if it had the original step length. These records keep a refinement experiment consistent with the fixed-time comparison in the theory.
Variable steps and a useful uniform bound
An equally spaced grid is convenient, rather than essential. Let the positive step lengths be , with their sum equal to , and let be their maximum. Euler uses the actual at each step. Under the same hypotheses, the defect is bounded by . Iterating the error inequality produces products of factors . Each product is at most , and . Therefore the simple estimate holds with exact initial data.
This bound is less sharp than the geometric-sum estimate on a uniform grid, but it proves a useful fact: convergence follows when the largest step tends to zero, even if the grid points are uneven. Having many very small steps does not compensate for retaining a single large step. Also, selecting a step from a local defect estimate does not by itself control the global error; the propagation bound remains part of the argument.
One final caution concerns the word order. An order-one guarantee allows special problems to be solved more accurately. For constant , Euler reproduces the exact linear solution at all grid points. For other problems the leading error may vanish at a particular final time. Order describes a guaranteed rate under stated hypotheses and, when called exact order, a rate attained for suitable problems. It is not a promise that every computed error halves precisely whenever the mesh is halved.
Common mistakes
Common mistake
(Evaluating the defect along the computed solution)
Consistency requires the equation to be evaluated along the exact solution. Replacing by in the defect changes the object into a numerical residual or error expression, which no Taylor argument controls.
Common mistake
(Calling Euler second order because its defect is )
The unscaled defect is and the scaled defect is . The order of the method is the order of the global error, which matches : about defects of size accumulate to , as the exact example shows with .
Common mistake
(Reading a fixed-time bound as a long-time guarantee)
A convergence theorem for fixed is not uniform as . For the bound grows like although both the solution and the actual error decay. Decay, oscillation, conservation and long-time behavior require sharper stability analysis, which motivates the next unit.
A Taylor argument also needs smoothness of the solution on the whole interval, not merely a formula that looks differentiable on part of the domain.
Counterexample mode
Without the Lipschitz condition the error need not vanish
One might expect continuity of and a smooth solution to be enough. Take , , on . The function is a solution, since ; it lies in with . Euler, however, computes at every step, so for every . The same initial value has the solutions for , for , one for each , so there is no unique solution to approximate. The state function is not Lipschitz near zero: as . With a Lipschitz constant restored, uniqueness and the error bound return.
Summary
Euler follows the vector field at the left endpoint. With bounded second derivative, one exact step has a defect of size at most . A Lipschitz bound controls how previous error propagates and yields a uniform error of order on a fixed time interval. A useful error statement names the norm, initial-data accuracy, smoothness assumptions, time interval and defect convention.
Exercises
Quick check
(Apply two Euler steps of size to , . Find and the signed error at , given .)
Evaluate each slope at the current time and computed state.
Solution · Answer
At the slope is , so . At the slope is , so . The exact value is , giving . Using the new state together with the old time would define a different method.
Quick check
(For , , compute the exact one-step defect at and show directly that the final error is of order on .)
Use and the sum of the first squared integers.
Solution · Answer
The defect is . For , summation gives . Hence . The coefficient depends on the varying local defects; the global error is first order. Since and , the general theorem also gives the valid bound .
Quick check
(Suppose , , , initial error is at most , and . Give a global error bound including both contributions.)
Use the exponential upper bound and retain the amplified initial error.
Solution · Answer
The bound is . This is an absolute upper bound in the chosen norm. Omitting the first term would assume exact initial data, which the question does not supply.