Can a deterministic calculation have a probability distribution?

The exact answer can be fixed while our knowledge of it is uncertain. The distribution represents that knowledge under the method's assumptions.

Does the uncertainty include bad data?

Not automatically. Measurement error and other sources need explicit treatment if they are to be included.

Is a probabilistic method always more accurate?

No. Performance depends on the problem, assumptions, calibration and computation budget. Conventional numerical methods remain important comparisons.

AI Glossary / AI and machine learning / Advanced

Probabilistic numerics

Definition

Probabilistic numerics treats numerical computation as inference about an unknown mathematical quantity. Instead of returning only an approximate answer, a probabilistic numerical method also represents uncertainty arising from the limited computational information used to obtain that answer.

What it means

An integral, equation solution or matrix calculation may be deterministic, yet we may not know its exact value because we can afford only a limited number of evaluations or iterations. Probability can describe uncertainty about the answer given that incomplete information.

This numerical uncertainty differs from randomness in the physical world and uncertainty about a scientific model. A calculation can quantify its own approximation uncertainty while still solving the wrong model. Keeping these sources separate helps avoid interpreting one uncertainty estimate as a complete account of everything unknown.

How it works

The method defines a probabilistic model for an unknown object, such as a function or equation solution. Computed observations update that model, producing a distribution for the desired numerical quantity. Further evaluations can be selected to reduce uncertainty where they are most informative.

Examples include probabilistic integration and approaches to differential equations or linear algebra. The usefulness of the reported uncertainty depends on modeling assumptions and calibration. A narrow distribution is not automatically trustworthy merely because it comes from a Bayesian calculation. Compare uncertainty estimates with known solutions or reliable reference computations where possible.

Three practical examples

1. A research team approximating an expensive integral can use probabilistic integration to express uncertainty due to observing the integrand at only selected points.

2. A simulation study can investigate how uncertainty from an approximate differential-equation solution influences later calculations.

3. A numerical-methods class can compare a single approximate answer with a distribution over possible answers, then test whether additional computation reduces uncertainty appropriately. These are illustrations of numerical inference, not automatic guarantees of scientific accuracy.

Think of it this way

Imagine estimating the shape of a road after surveying only a few locations. You can draw an estimated route and also indicate where its position remains uncertain. Another survey measurement can improve both the estimate and your confidence.

The analogy concerns incomplete knowledge. It does not mean that the underlying road, or a deterministic integral, changes randomly each time you look at it.

Worked example

Suppose a hypothetical probabilistic integration method returns an approximate Gaussian distribution for an integral, with mean 10 and standard deviation 0.4. Under that stated approximation, a central interval of roughly two standard deviations extends from 9.2 to 10.8.

After additional evaluations, suppose the mean becomes 10.1 and the standard deviation becomes 0.1, giving a corresponding approximate interval from 9.9 to 10.3. The numerical estimate has changed and its modeled uncertainty has narrowed.

These invented outputs illustrate interpretation, not a complete integration algorithm. Whether the interval has useful coverage must be assessed. It does not include uncertainty from an incorrect integrand model or inaccurate physical input data unless those have been explicitly incorporated.

Illustrative case study

Fictional simulation team NumerisLab estimates a quantity used in an engineering comparison. Its ordinary numerical routine returns one value, making it difficult to tell whether the small difference between two designs exceeds computational approximation error.

The team investigates a probabilistic numerical method and checks its uncertainty estimates on simpler cases with reference answers. Some initial intervals are too narrow. Researchers revise the assumptions and compare the result with conventional error controls. The exercise helps identify when more computation is needed; it does not make the underlying engineering model automatically correct.

Common misconceptions and mistakes

Probabilistic numerics does not mean that every numerical problem is inherently random. The probability distribution can describe limited knowledge of a deterministic answer. It is also not synonymous with Monte Carlo simply because both use probability-related ideas.

Another mistake is combining numerical, measurement and model uncertainty without stating how they relate. A solver's uncertainty can be small while the scientific model is inaccurate. Avoid treating posterior spread as a validated error bar without calibration evidence. Additional computation may reduce one uncertainty source while leaving other sources almost unchanged.

Business use

This field matters when expensive numerical calculations influence decisions and approximation error is material. It can support questions about whether extra computational effort is worthwhile and how uncertainty moves through a sequence of calculations.

A project proposal should identify the numerical quantity, the information being computed and the uncertainty the method claims to quantify. Compare it with established error-estimation methods at a similar cost. For stakeholders, report the practical consequence: whether the remaining numerical uncertainty could change the decision. The probabilistic formulation is a means to that end, not a business outcome by itself.

Check your understanding

A numerical solver reports very small uncertainty, but its physical model omits an important effect. Does the small uncertainty establish an accurate real-world prediction?

Answer: No. Numerical uncertainty concerns the calculation being performed. An incomplete physical model can still produce a precisely calculated but misleading prediction.

Related terms

Gaussian process
Calibration
Monte Carlo integration
Numerical analysis
Bayesian quadrature
uncertainty quantification
differential equation solver
approximation error

Learn next

Separate measurement, model and numerical uncertainty in one simulation workflow. Then compare an approximate result with a known answer while varying computational effort. Check both the estimate and whether reported uncertainty reflects the observed errors.

Measurement model

Sources

Hennig, Osborne and Girolami, Probabilistic Numerics and Uncertainty in Computations: https://arxiv.org/abs/1506.01326. Probabilistic Numerics research workshop overview: https://www.probabilistic-numerics.org/meetings/2015_Warwick/. The illustrative intervals and NumerisLab scenario are original; they are not measured results from those sources. Independent review consulted the full primary paper PDF, including its distinction between numerical epistemic uncertainty and uncertainty in the problem specification, and its discussion of uncertainty calibration: https://arxiv.org/pdf/1506.01326. The illustrative intervals were checked independently.

Previous: Probabilistic model Next: Probabilistic principal component analysis