Access Official Technical PDF

Share Summary:LinkedInEmail

Computational accuracy matters wherever engineering calculations inform industrial automation decisions. A computer can produce results with many displayed digits, but those digits do not guarantee that the result is reliable. Input uncertainty, finite numerical precision, rounding and the sequence of calculations can all affect what a number means. For engineers and system integrators, the practical task is to distinguish useful precision from spurious accuracy—and to make that distinction clear in calculations, models and reported results.

Why computational accuracy needs scrutiny

Fast computation can encourage confidence in an answer simply because it was produced quickly and expressed precisely. Yet a result may be affected by errors that are not obvious from its appearance. It is important to separate expected or tolerated errors from unexpected or overlooked ones. A calculation may be internally consistent and still be unsuitable for a decision if its assumptions, limits or numerical behavior have not been considered.

This is especially relevant when calculations are repeated, when results feed later calculations, or when a model represents physical behavior. The number of digits printed is not, by itself, evidence that those digits carry meaningful information. Precision should be assessed in relation to the input data, the calculation method and the tolerance the application requires.

Floating-point arithmetic and machine limits

Computers represent many values using floating-point arithmetic: a fixed-length significand combined with an exponent. The available range and precision depend on the computer’s representation and implementation. Because the representation is finite, some operations require rounding, and not every mathematically possible value can be represented exactly.

Several conditions deserve attention in numerical software:

  • Rounding: Small differences introduced during arithmetic can affect later operations, particularly across a long sequence of calculations.
  • Overflow and underflow: A value may exceed the representable range or become too small for the system’s representation. Programs and system settings need a defined response to these conditions.
  • Zero and near-zero values: A computed value close to zero is not necessarily equivalent to exact zero. Dividing by a value that is zero—or treated as zero under a tolerance—can create problems.
  • Variable limits: Applying expected limits to variables can help a program detect an out-of-range result earlier, rather than allowing it to affect subsequent calculations unnoticed.

These checks do not eliminate numerical error. They make important conditions more visible and help ensure that exceptional values are handled deliberately.

Curve fitting: more complexity is not always better

Empirical data is often represented by a fitted curve or polynomial rather than stored as a table. A fitted expression can be convenient, but increasing its degree does not automatically improve the result. Higher-order terms may make a fit worse, and a single expression may be unsuitable across a range where the underlying behavior changes.

For practical modeling, dividing the range into sections may be preferable to using an increasingly high-degree polynomial. The choice should be guided by the required tolerance and the behavior of the data, not simply by the availability of a more complex calculation. Coefficients also warrant care: carrying unnecessary digits can create a misleading impression of certainty. Simplifying expressions and reducing avoidable computation can help limit opportunities for error.

Track error through successive calculations

Errors can accumulate or interact as a result passes through multiple operations. A detailed analysis of coincidental errors may be impractical in some cases, but teams can still compare calculations made under different numerical conditions. For example, comparing rounded and unrounded results can reveal sensitivity to rounding. Repeating a calculation independently or checking a result through reinversion may also help identify discrepancies.

Intermediate calculations may use more precision than the final reported value requires. However, retaining extra digits internally should not be confused with having greater certainty in the result. At the end of the calculation, reported precision should reflect the accuracy the method and inputs can support. Where results vary with parameter combinations, presenting a range or tolerance can communicate that uncertainty more honestly than reporting an apparently exact value.

Step size and accuracy in numerical methods

When numerical methods approximate the solution of differential equations, step size affects both the calculation and its workload. Variable-step approaches, including methods associated with Adams and Runge–Kutta, adjust the step according to the error observed or allowed. A practical strategy described in the source material is to halve the step when error exceeds a permitted maximum and double it when the error falls below a minimum requirement.

Behavior near a singularity or other difficult region may require particular attention. Repeatedly reducing the step cannot continue indefinitely: the computer has a smallest representable increment for the calculation. Programs therefore need sensible limits and a way to detect when further refinement is no longer possible or useful.

Put numerical results in operational context

For industrial automation teams, the key lesson is to treat computed output as an engineering result that needs context—not as a collection of self-validating digits. Define acceptable tolerances, check variable ranges, account for finite precision and examine how errors may propagate through successive calculations. When using fitted data or numerical integration, select methods and settings that match the behavior being modeled and the accuracy required.

Finally, present results with a precision and tolerance that users can interpret. These practices do not remove every source of uncertainty, but they make calculations easier to assess and reduce the risk of mistaking numerical detail for dependable accuracy.

Access Official Technical PDF

About The Author