What Does It Mean for an Euler Approximation to Be Reliable?
An Euler computation produces approximate values for a solution of an initial-value problem. A small-looking number of decimal places does not, by itself, establish that those values are close to the true solution. Reliability is a judgment supported by several kinds of evidence: comparison with a refined computation, an error estimate where its assumptions hold, a stability check, and consistency with known qualitative behavior.
These checks answer different questions. Refinement asks whether the numerical answer changes as the step size decreases. An error bound can give a guaranteed limit on the discrepancy, but only under stated assumptions. Stability concerns how errors and perturbations behave during the computation; stability alone does not ensure accuracy. Qualitative checks look for behavior the model or differential equation requires, such as positivity or approach to an equilibrium.
A Refinement Comparison
For the initial-value problem \(y'=f(t,y)\), \(y(t_0)=y_0\), forward Euler with step size \(h\) uses the recurrence \(y_{n+1}=y_n+h f(t_n,y_n)\), where \(t_{n+1}=t_n+h\). To compare step sizes fairly, calculate to the same final time from the same initial data using the same equation. If the interval is fixed, halving \(h\) doubles the number of steps.
Euler’s method is a first-order method: under suitable smoothness and boundedness conditions, its global error is typically proportional to \(h\) when \(h\) is sufficiently small. Thus, if \(Y_h\) and \(Y_{h/2}\) are approximations at the same endpoint, the difference \(|Y_{h/2}-Y_h|\) can estimate the error in the finer result. In the asymptotic regime, the coarse error is about twice the fine error, so their difference is about the fine error. This is an estimate, not a universal guarantee: it may be misleading when steps are too large, the solution changes sharply, or the method is outside the regime where the error behaves proportionally to \(h\).
Worked Example: Refining Euler’s Step for Exponential Decay
Consider \(y'=-y\), \(y(0)=1\), and estimate \(y(1)\). The exact solution is \(y(t)=e^{-t}\), so \(y(1)=e^{-1}\approx0.367879\). First use \(h=0.5\). The Euler multiplier at each step is \(1-h=0.5\), and there are two steps:
Now use \(h=0.25\). There are four steps, and each value is \(0.75\) times the preceding value:
The refinement difference is \(|0.31640625-0.25|=0.06640625\). Since Euler is first order, this difference gives a rough estimate of the error in the finer approximation when the computation is in the asymptotic regime. Checking against the exact value, the fine-grid error is \(|0.31640625-0.367879|\approx0.051473\), while the coarse-grid error is \(|0.25-0.367879|\approx0.117879\). The refinement comparison tracks the error’s scale but is not exact. Substitution into the Euler recurrence verifies both computed endpoints.
A Conditional Error Bound
Refinement can be supplemented by a rigorous bound when the solution and the differential equation satisfy useful conditions. Suppose the exact solution \(y(t)\) exists on \([t_0,T]\), is twice differentiable, and \(|y''(t)|\le M\) throughout that interval. Suppose also that \(f\) is Lipschitz in its second argument there: for some \(L\ge0\), \(|f(t,u)-f(t,v)|\le L|u-v|\) for all relevant \(u,v\). Starting from the exact initial value, let \(e_n=|y(t_n)-y_n|\) denote the error at a grid point, where \(t_n=t_0+nh\).
Here is the argument. Taylor’s theorem with a bounded second derivative gives \(y(t_{n+1})=y(t_n)+h y'(t_n)+r_n\), where \(|r_n|\le Mh^2/2\). Since \(y'(t_n)=f(t_n,y(t_n))\), compare this exact step with the Euler step and use the Lipschitz condition:
The initial value is exact, so \(e_0=0\). Repeatedly applying the recurrence adds a geometric sum:
This proves the first bound. The inequality \(1+x\le e^x\) for \(x\ge0\) gives the exponential form. If \(L=0\), the recurrence becomes \(e_{n+1}\le e_n+Mh^2/2\), and adding this inequality \(n\) times gives \(e_n\le nMh^2/2\). The bound can be conservative, and finding valid \(M\) and \(L\) may be difficult. It applies only when the stated conditions hold over the interval.
Worked Example: Applying a Global Error Bound
For \(y'=-y\), \(y(0)=1\), compute to \(T=0.5\) using \(h=0.1\). The exact solution is \(e^{-t}\), so \(y''(t)=e^{-t}\), and \(0<e^{-t}\le1\) on \([0,0.5]\). We may take \(M=1\). Also, \(|f(t,u)-f(t,v)|=|-u+v|=|u-v|\), so \(L=1\). There are \(n=5\) steps. The discrete error bound is therefore
The Euler calculation gives \(y_5=0.9^5=0.59049\). The exact endpoint is \(e^{-0.5}\approx0.60653066\), so the actual error is \(|0.59049-0.60653066|\approx0.01604066\), which is below the bound. The bound is valid but not sharp. The recurrence \(y_{n+1}=0.9y_n\) verifies the computed Euler endpoint.
Stability and Qualitative Behavior
A stable computation can still be inaccurate. For the scalar test equation \(y'=\lambda y\), one Euler step multiplies the current value by \(1+h\lambda\). When \(\lambda<0\), the exact solution decays. The Euler values do not grow in magnitude from one step to the next if \(|1+h\lambda|\le1\); however, this condition does not guarantee that the numerical values closely approximate the exact solution. For instance, a multiplier of zero prevents growth but can erase a nonzero solution too quickly.
Qualitative checks provide another independent test. Depending on the problem, the solution may be expected to stay positive, move in a particular direction, approach an equilibrium, or avoid a singularity. A numerical result that violates a well-established feature may signal an unsuitable step size, an implementation error, or a model issue. The check must be appropriate to the differential equation: a decreasing quantity should not be assumed positive unless the equation and initial data support that conclusion.
Worked Example: Stability Does Not Guarantee Accuracy
Consider \(y'=-2y\), \(y(0)=1\), and estimate \(y(1)\). With \(h=0.5\), the Euler multiplier is \(1-2(0.5)=0\). Two steps give \(y_1=1+0.5(-2) =0\) and \(y_2=0+0.5(0)=0\). The stability factor has magnitude zero, so the computed values do not grow, but the result is not accurate: the exact value is \(e^{-2}\approx0.135335\).
Refine to \(h=0.25\). The multiplier is \(1-2(0.25)=0.5\), and four steps give \(1,\ 0.5,\ 0.25,\ 0.125,\ 0.0625\). Thus the refined endpoint is \(0.0625\), still below the exact value. The finer result is positive and decreasing, as expected, and is closer to the exact endpoint than the coarse value. Substitution into \(y_{n+1}=(1-2h)y_n\) verifies each grid sequence. The comparison supports refinement, but also shows that stability and qualitative agreement alone do not certify a chosen accuracy tolerance.
Using Several Checks Together
A defensible reliability judgment combines evidence rather than relying on one diagnostic. A practical sequence is:
Specify the endpoint or quantity being estimated and the accuracy required. Compare approximations at the same target using the same initial data and model.
Compare the result for \(h\) with the result for \(h/2\), and, when feasible, refine again. Look for changes that become smaller in a consistent way.
Use a justified error bound if its hypotheses and constants are known. Otherwise label a refinement-based estimate as an estimate, not a guarantee.
Look for step-size-driven growth or oscillation and compare the computed trajectory with qualitative properties supported by the equation and initial data.
Report what the checks support, the assumptions behind them, and what remains unverified. Do not claim more accuracy than the evidence establishes.
A Common Mistake: Treating Agreement as Proof
A frequent error is to say that a refined answer must be accurate because two step sizes give similar values. Small differences can be reassuring, but both computations can share a systematic error, or the step sizes can still be too large to reveal rapid changes. Refinement is more informative when several successive comparisons show a consistent trend and when independent error or qualitative checks agree.
For example, if a coarse run gives \(0.500\) and a refined run gives \(0.499\), the difference is \(0.001\). It is incorrect to conclude, without additional assumptions, that the refined answer’s error is at most \(0.001\). The true answer could differ from both values. The correction is to state that the two results differ by \(0.001\), then identify whether a valid error bound, a justified asymptotic estimate, or other evidence supports a stronger claim.
What the Evidence Can Establish
Euler reliability is not a property conveyed by the number of displayed digits. It depends on the question being asked, the step size, the behavior of the differential equation, and the quality of the available checks. A refinement study can reveal sensitivity to step size. A theorem-based bound can guarantee an error limit when its assumptions and constants are valid. Stability and qualitative checks can expose implausible behavior. None should be mistaken for another.
When no rigorous bound is available, a careful conclusion can still be useful: state the values obtained at successive step sizes, describe whether the results appear to converge, note any stability or qualitative concerns, and say explicitly that the numerical evidence does not prove a particular error tolerance. This distinction between evidence and guarantee is essential whenever a numerical solution informs further calculation or interpretation.
Check Your Understanding
For each question, distinguish a numerical observation from a guaranteed conclusion.
- What must be held fixed when comparing Euler approximations with step sizes \(h\) and \(h/2\)?
- Why can the difference between two refined Euler results estimate an error without guaranteeing its size?
- What conditions are needed for the global error bound proved above?
- For \(y'=\lambda y\), what does \(|1+h\lambda|\le1\) say, and what does it not say?
- Give one reason why agreement between two step sizes may still provide insufficient evidence of reliability.
