Does the method assume normality?

No. Normality can make its formulas coincide with full Bayesian conditioning, but is not required for the linear adjustment itself.

Is it ordinary linear regression?

No. It revises specified beliefs about uncertain quantities; similarities in formulas do not make the frameworks identical.

Can observing D leave B unchanged?

Yes, for example when their specified covariance is zero, although that does not prove independence.

AI Glossary / Statistics and probability / Advanced

Bayes linear statistics

Definition

Bayes linear statistics is an approach to representing and revising uncertainty using expectations, variances, and covariances rather than requiring a fully specified joint probability distribution. It updates these second-order beliefs through linear adjustment when observations become available. The method is associated particularly with Michael Goldstein, David Wooff, and their collaborators.

What it means

A full Bayesian analysis can require probability distributions for many uncertain quantities and their relationships. In a complicated system, specifying that much detail may be difficult to justify or computationally burdensome.

Bayes linear analysis begins with a more limited description: what values are expected, how uncertain they are, and how the quantities vary together. Observations then revise the expectations and associated residual uncertainty through the specified covariance structure.

This is a distinct framework, not simply another name for Bayesian linear regression. Linear describes the adjustment in the observed quantities. It does not require the physical system itself to obey a straight-line law, although the chosen belief structure limits the relationships that the adjustment can exploit.

How it works

Let B be an uncertain quantity or vector of interest and D the quantities to be observed. Specify their expectations, variance matrices, and cross-covariance coherently, with finite second moments.

When D takes the observed value d, the adjusted expectation is E(B) + Cov(B,D) Var(D) inverse times [d - E(D)]. It is the affine function of D that minimizes expected squared discrepancy under the specified second-order beliefs.

The adjusted variance is Var(B) - Cov(B,D) Var(D) inverse Cov(D,B). It describes residual variation after the linear adjustment. A suitable generalized inverse can handle a singular data variance matrix, with attention to consistency and numerical stability.

These expressions resemble conditional mean and variance formulas for a joint Gaussian distribution. In that special setting they coincide with those full conditional quantities. Generally, a Bayes linear adjusted expectation is not the full conditional expectation, and adjusted variance is not automatically a data-dependent conditional variance. In the basic fixed-covariance adjustment, the variance reduction depends on the specification and observed variables, not on their realized numerical values.

Three practical examples

Computer experiments: update beliefs about an expensive simulator's outputs using a limited set of runs and an explicitly chosen covariance structure.
Inspection planning: examine how observations of selected components could reduce uncertainty about related unobserved components.
Complex forecasting: combine partial expert judgments and observations when a coherent second-order specification is more practical than a complete probabilistic model.

Think of it this way

Think of connected measuring gauges. A reading above expectation on one gauge changes expectations for another according to how strongly the gauges are believed to move together. Bayes linear adjustment formalizes that revision without requiring a complete probability story for every possible combination of readings.

Worked example

Suppose an uncertain quantity B has prior expectation 100 and variance 25. An observable indicator D has expectation 50 and variance 9. Their covariance is 12. These numbers are coherent: the covariance matrix determinant is 25 × 9 - 12 squared = 81, which is positive.

Now observe D = 56, six units above its expected value. The adjustment coefficient is 12 / 9 = 4/3. The adjusted expectation of B is 100 + (4/3) × 6 = 108.

The adjusted variance is 25 - (12 squared / 9) = 9, so the residual standard deviation is 3. The adjustment resolves 16 of the original 25 variance units, a fraction of 16/25 = 64%.

If the observed indicator had instead been 44, the adjusted expectation would be 92, while this basic adjustment would still yield variance 9. These moment calculations alone do not justify attaching a normal 95% probability interval to 108. That interpretation would require additional distributional assumptions or a separately justified uncertainty statement.

Illustrative case study

Fictional case: Rowan Engineering runs a costly simulation of a cooling system. Analysts can specify expected outputs and plausible covariance patterns across operating settings more confidently than they can justify an entire joint distribution.

They construct a Bayes linear emulator and use selected simulator runs to adjust predictions for untested settings. Before using the emulator to screen designs, they compare predictions with additional reserved runs and examine unexpectedly large discrepancies.

The team revises an implausibly strong covariance assumption when it overstates how much one region informs another. Managers receive adjusted expectations and uncertainty measures with the specification recorded, rather than a claim that limited moment assumptions provide a complete posterior distribution.

Common misconceptions and mistakes

Do not treat the method as assumption-free. Expectations and covariances are substantive judgments that can be misspecified, and the joint covariance structure must be valid.

Do not assume zero covariance means independence. The basic linear adjustment may find no useful linear relationship even when nonlinear dependence exists.

Do not automatically label the adjusted variance a posterior variance. The two coincide under additional conditions, but they are not interchangeable in general.

Do not ignore surprising observations merely because the adjustment formula is easy to calculate. Diagnostics can reveal tension with the original beliefs. Covariance choices also control apparent uncertainty reduction, so an overconfident specification can produce misleadingly precise results.

Business use

The business value is a tractable way to learn from interconnected uncertain quantities when full probabilistic modeling would require unsupported detail. It can also support decisions about which observations are expected to be informative.

Teams should document how expectations and covariances were obtained, examine sensitivity, and validate predictions where observations are available. If decisions depend strongly on tail probabilities, hard constraints, or nonlinear interactions, second-order summaries alone may be insufficient. The choice between Bayes linear and fuller probabilistic analysis should follow the information and decisions required.

Check your understanding

Why does the worked example move the expectation upward?

Answer: The observed indicator exceeds its expectation and has positive covariance with the target.

Does adjusted variance 9 by itself imply a normal posterior?

Answer: No. A variance specifies only part of the uncertainty description.

Related terms

Covariance
Bayesian inference
Gaussian process
Linear prediction
Adjusted expectation
second-order beliefs
emulator
uncertainty quantification

Learn next

Explore how different covariance assumptions change the adjustment and variance reduction. Then study diagnostic checks and the conditions under which Bayes linear adjustment agrees with full probabilistic conditioning.

Covariance

Sources

Michael Goldstein, Bayes Linear Methods I, author-hosted introduction and sections on adjusted expectation and variance: https://maths.dur.ac.uk/stats/bd/blm1.online/blm1.html Multi-Output GP Emulator documentation, Calculation of adjusted expectation and variance, explicit update formulas: https://mogp-emulator.readthedocs.io/en/latest/methods/proc/ProcBLAdjust.html

Previous: Bayes factor Next: Bayes-Nash equilibrium