1. Definition and Basic Concepts
Global error is the difference between an exact quantity of interest and the approximation produced by a numerical procedure applied over an interval. The “exact” target may be a function value, an invariant, or the value of a solution to a differential or functional problem. In practice, global error is contrasted with local error, which describes what goes wrong on a single step or local region.
1.1 Error decomposition: local vs. global error
Local error typically measures the discrepancy introduced by one step of a method, often expressed through the method’s local truncation error. Global error measures the cumulative effect of these local discrepancies across all steps, combined with how earlier mistakes influence later computations. As a result, global error is usually not a simple sum of local errors; it is affected by propagation through the algorithm.
1.2 Error norms and how “size” of error is measured
To quantify global error, one chooses a norm that maps an error function (or error vector at a grid of points) to a scalar. Common choices include norms in which the maximum magnitude dominates (useful for worst-case behavior), norms that aggregate magnitude across points (useful for overall accuracy), or energy-like norms when the problem has a natural structure. The selected norm influences both theoretical bounds and how performance is reported.
1.3 Absolute error versus relative error
Absolute error compares an approximation to the true value using subtraction or an equivalent measure of distance. Relative error scales the difference by the magnitude of the true quantity, which is helpful when the target varies widely. Relative measures can become unstable near zero, so practitioners often use safeguards such as mixed criteria (e.g., absolute plus relative tolerance).
1.4 Pointwise error versus error over an interval
Pointwise error is evaluated at specific points, such as grid nodes or evaluation times. Error over an interval describes accuracy of the entire computed trajectory or function approximation on a continuous region, typically through a norm or integral measure. While pointwise views are convenient for diagnostics, interval-level measures better reflect how the numerical method behaves throughout the computation.
2. Relation to Numerical Methods
Global error is tightly linked to how a numerical method advances or builds approximations. Many methods can be viewed as repeatedly applying an update rule; global error then reflects how perturbations injected at each update propagate to the final output.
2.1 Step-by-step accumulation of error
In iterative or time-stepping algorithms, each step uses current information to predict the next state. If a step introduces an error, later steps consume that imperfect state. Under mild conditions, the accumulated effect often grows in a structured manner with the number of steps and the problem’s dynamics.
2.2 Global error in one-step methods
One-step methods (typical examples include standard Runge–Kutta schemes) update the solution using information from a single previous point. For such methods, global error often inherits a predictable relationship with the method’s order and the step size. Conceptually, the method’s local truncation error determines how strongly each step injects error, while the one-step structure governs how that error transfers forward.
2.3 Global error in multi-step methods
Multi-step methods use several previous states. This increases the ways errors can enter: not only through the most recent update, but through the memory of earlier approximations. Global error analysis may therefore involve both the accumulation of new local errors and the persistence of past inaccuracies, sometimes requiring additional assumptions on stability and consistency.
2.4 Stability effects on global error growth
Stability describes whether small perturbations in the input or intermediate values remain small as the computation continues. An unstable method can amplify local inaccuracies dramatically, causing global error to grow faster than what order alone would suggest. Stable methods may still accumulate error, but in a controlled fashion aligned with theoretical expectations.
2.5 Consistency, stability, and convergence (overview)
In numerical analysis, convergence is the property that the approximation approaches the exact solution as step size tends to zero. Consistency relates to how well the method reproduces the underlying problem in the limit of small steps, while stability controls error propagation. Together, these ideas explain why global error typically decreases as the discretization is refined, and why certain methods fail despite good local behavior.
3. Connection to Truncation Error and Convergence Order
Global error is frequently connected to truncation error through the method’s order, which captures the dominant rate at which errors vanish as the step size shrinks.
3.1 Local truncation error (LTE) recap
Local truncation error quantifies the defect when the exact solution is substituted into the numerical update rule over a single step. For a method with a certain order, the LTE typically scales like a power of the step size, often expressed as \(O(h^{p+1})\), where \(h\) is the step size and \(p\) is the convergence order in many common settings.
3.2 From LTE to global error: typical scaling
Under standard assumptions, the number of steps grows like \(1/h\) over a fixed time or interval length. If local defects behave like \(h^{p+1}\), their cumulative effect often results in global error scaling like \(h^{p}\). This “one less power” relationship is a common pattern, though detailed behavior can depend on stability and problem regularity.
3.3 Convergence order and practical interpretation
| Convergence order indicates how rapidly global error decreases when step size is reduced. For example, an order-\(p\) method often satisfies a relation of the form \(\|e(h)\| \approx C h^{p}\) for sufficiently small \(h\), where \(e(h)\) is the global error and \(C\) is a constant reflecting problem and method details. In practice, this guides expectations for accuracy improvements under refinement. |
|---|
3.4 Empirical order of convergence (EOC)
Empirical order is estimated from computed errors at several step sizes, typically using differences between solutions obtained at different refinements. If \(E(h)\) denotes an observed error measure, EOC is inferred from how \(E\) scales with \(h\). EOC is valuable when exact errors are unknown and when theoretical constants are difficult to assess.
3.5 Impact of step size on global error
Step size controls both truncation and accumulation effects. Too large a step increases truncation error, typically dominating the global error. Too small a step can increase the relative influence of round-off error and amplify sensitivity to floating-point limitations. The total global error often has an “optimal region” where decreasing \(h\) yields real gains before numerical noise takes over.
4. Global Error Analysis in Common Calculus Settings
Global error appears across calculus-oriented numerical tasks, where exact analytic quantities are replaced by discrete or approximate computations.
4.1 Approximating definite integrals (quadrature errors)
When computing definite integrals numerically, quadrature rules approximate the integral using weighted samples. The global quadrature error refers to the difference between the exact integral and the computed value after applying a chosen rule or composite scheme over the interval. Its magnitude depends on smoothness of the integrand, the degree of exactness of the rule, and how the interval is partitioned.
4.2 Derivative approximations and discretization error
Derivatives are often approximated using finite differences or spectral-like discretizations. The resulting discretization error is global in the sense that it affects derivative values used throughout subsequent computations (such as in solving differential equations). Depending on the stencil and boundary treatment, error may vary across the domain, making both pointwise and norm-based measures relevant.
4.3 Error in series truncation versus discretization
Series-based computations introduce an error by truncating an infinite expansion, while discretization introduces error by approximating continuous operators on finite grids. These two sources can coexist. For instance, in a method that uses an expansion inside a discretization scheme, truncation error may dominate for moderate resolution, while finer discretizations reduce grid error until truncation becomes the limiting factor.
4.4 Error propagation through function composition
Many calculus computations involve nested approximations, such as evaluating a composition \(g(f(x))\) where both \(f\) and \(g\) are approximated. Errors in the inner approximation can be transformed or amplified by the outer function, often depending on derivatives (or sensitivities) of the outer map. As a result, global error is not merely the sum of component errors; it reflects how each stage modulates uncertainty from earlier stages.
5. Bounds, Estimates, and Theoretical Results
Theoretical tools aim to bound global error without computing it exactly. These results clarify how error depends on step size, method order, and stability properties.
5.1 A priori error estimates
A priori estimates bound global error using information available before the computation, such as step size, method parameters, and assumptions on the exact solution’s smoothness. Such bounds often express the asymptotic rate and sometimes an upper constant, yielding guidance on what accuracy should be achievable when choosing discretization parameters.
5.2 A posteriori error estimates
A posteriori estimates use information from the computed solution itself, frequently incorporating residuals or differences between solutions on related meshes. They can provide more reliable indicators when assumptions for a priori theory are too conservative. In practice, these estimates support adaptive refinement strategies by identifying where and how strongly the approximation fails.
5.3 Grönwall-type bounds (high-level usage)
Grönwall-type inequalities are common in controlling how errors evolve in time-dependent problems. At a high level, they convert a differential or integral inequality describing error growth into an explicit bound, often leading to exponentials of accumulated quantities. While the resulting expressions can be pessimistic, they explain why stability and problem growth rates matter for global error.
5.4 Worst-case bounds versus average behavior
Theoretical bounds often represent worst-case scenarios under broad assumptions, which can overestimate observed errors. Empirical performance may be better because error signs partially cancel, the solution behaves more regularly than assumed, or perturbations are localized. Consequently, practitioners may rely on bounds for safety and use diagnostics for refinement rather than treating bounds as exact predictions.
5.5 Round-off error versus truncation error (separation concept)
Global error comprises at least two major contributors: discretization (or truncation) error and floating-point round-off error. Separation concepts in analysis treat these contributions differently: truncation tends to decrease with step size, while round-off tends to increase as operations multiply. Understanding this distinction helps explain the existence of regimes where refining the grid improves accuracy and regimes where it degrades it.
6. Numerical Experiments and Diagnostics
Because exact global error is usually unknown, experiments and diagnostic tools are central to understanding accuracy and identifying limiting factors.
6.1 Refinement studies and error plots
A refinement study computes approximations with decreasing step sizes and compares them using error surrogates (such as differences between successive refinements). Plotting these measures versus step size typically reveals whether the method’s convergence rate matches expectations. Deviations can signal stability problems, insufficient smoothness, or contamination by round-off.
6.2 Choosing step sizes to meet tolerance
When an accuracy target is specified, step size must be selected so the predicted global error falls below a tolerance. This may use heuristic rules based on observed convergence, or it may rely on error estimators in adaptive algorithms. In either case, good practice includes accounting for both truncation and round-off risks, especially when tolerances are set very tight.
6.3 Richardson extrapolation and error reduction (conceptual use)
Richardson extrapolation improves accuracy by combining results from multiple discretizations. If the error behaves like a power series in the step size, then suitable linear combinations can cancel leading-order terms. While the technique is conceptually simple, its effectiveness depends on having sufficiently small step sizes and consistent error models.
6.4 Sensitivity to initial conditions and parameters
Even when a method is stable, the problem itself may amplify differences in initial data or parameters. In such cases, global error includes not only numerical defects but also how the dynamics magnify perturbations. Diagnostics often involve varying initial inputs within plausible ranges to determine whether observed discrepancies are numerical or intrinsic to parameter uncertainty.
6.5 Detecting error domination regimes (truncation vs. round-off)
A common diagnostic is to observe what happens to accuracy as step size decreases. If error decreases following an expected order trend, truncation error likely dominates. If further refinement stops improving results or makes them worse, round-off error and loss of significance are likely taking over. Identifying the crossover region supports choosing practical step sizes that balance competing error sources.
7. Practical Computation Guidelines
Effective use of numerical methods requires translating error theory into robust computational practice.
7.1 Error budgeting across stages of a computation
Complex computations often contain multiple error sources: quadrature error, discretization error, iteration stopping error, and rounding effects. Error budgeting allocates a total tolerance among stages so no single component silently dominates. This approach supports predictable overall accuracy and helps prevent wasted computation on overly accurate intermediate steps.
7.2 Using adaptive step sizes to control global error
Adaptive strategies adjust step size based on local information to drive an estimate of global (or cumulative) error toward a target. This is especially helpful when the solution changes rapidly in time or the integrand has localized features. Adaptive control aims to maintain accuracy while reducing unnecessary work in regions where the solution is smooth.
7.3 Guarding against overfitting accuracy (tol too strict)
Setting tolerances too aggressively can cause excessive computation and can also increase the influence of floating-point noise, yielding results that appear stable but are not genuinely more accurate. Guarding against this involves monitoring convergence behavior, checking EOC trends, and recognizing when further refinement fails to improve (or worsens) the computed quantity.
7.4 Software considerations and reproducibility
Implementation details such as stopping criteria, error estimator configuration, floating-point precision, and compiler settings affect global error in practice. Reproducibility requires recording these choices, along with step-size policies and random seeds if relevant. Versioning of numerical libraries is also important because small algorithmic changes can alter error behavior.
7.5 Reporting global error in computational work
Clear reporting helps interpret results and compare methods. Since exact global error is rarely known, authors typically report estimated errors, residual-based indicators, or convergence study outcomes, along with the norms used. Including both the achieved accuracy and the computational parameters that produced it enables readers to assess whether the method’s performance matches the expected global error behavior.