Why does the smaller variance get more weight?

Under the stated model, that estimate provides a more precise observation of the common quantity.

Can I use standard deviations directly as denominators?

That gives a different rule; inverse-variance weights use the squared standard deviations.

Does the formula handle correlated errors unchanged?

No. Dependence requires an appropriate covariance model for the same optimality and uncertainty claims.

AI Glossary / Computer vision and multimodal AI / Advanced

Inverse-variance weighting

Definition

Inverse-variance weighting combines estimates by assigning larger weights to those with smaller variances. For independent, unbiased estimates of the same quantity with known positive variances, normalized weights proportional to one divided by variance minimize the variance among linear unbiased weighted averages. The stated conditions are central to that optimality claim.

What it means

A less variable measurement receives more influence because it is more precise under the model. Precision is not the same as freedom from bias: a consistently wrong sensor can have a small variance. The method also assumes the estimates target a common quantity. Combining measurements of different underlying conditions merely because they use the same units can be misleading. Estimated variances introduce further uncertainty, so weights should not be treated as unquestionable facts when they come from limited or unrepresentative calibration data.

How it works

For estimates xi with variances vi, set raw weights wi=1/vi and compute the combined estimate as the sum of wi×xi divided by the sum of wi. Under independence, the combined variance is 1 divided by the sum of inverse variances. Correlated errors require covariance-aware treatment rather than the independent formula. The weights use inverse variance, not inverse standard deviation. Zero or nearly zero variance estimates also require careful interpretation because they can create extreme dominance without establishing that an observation is perfectly accurate.

Three practical examples

An imaging system combines independent estimates of the same intensity after placing them on a compatible scale and characterizing their noise.
A measurement team fuses calibrated sensor readings while checking whether shared environmental errors violate the independence assumption.
A researcher compares equal weighting with precision weighting in a controlled simulation where the common target and variances are known.

Think of it this way

Imagine asking two careful counters to estimate the same inventory total. One method varies only a little across repeated controlled trials, while the other varies more. Giving the steadier method more influence can reduce random variation in the combined estimate. However, if that counter always overlooks one shelf, steadiness does not establish correctness. The analogy separates precision from bias and explains why a weighting rule cannot replace checking that both methods measure the intended quantity.

Worked example

In an original independent-measurement example, estimate A is 10 with variance 4, and estimate B is 14 with variance 16. Their inverse-variance weights are 1/4 and 1/16, which normalize to 0.8 and 0.2. The combined estimate is 0.8×10+0.2×14=10.8. Its variance under the stated assumptions is 1/(1/4+1/16)=3.2. An equal average has variance (4+16)/4=5 under independence. The lower variance does not prove that 10.8 is closer to the unknown truth in this particular realization; it describes expected variability under the model.

Illustrative case study

A fictional imaging team combines repeated measurements from two acquisition settings. Initial weighting favors the less variable setting, but both share a drifting calibration offset. The team recognizes that the independent-error calculation understates uncertainty and separately investigates the shared error. It also checks whether the measurements were converted to the same scale before computing weights. This scenario illustrates why a mathematically correct formula can be misapplied to an unsuitable error model. It does not claim that correlation always makes fusion useless or that equal weighting automatically solves calibration bias.

Common misconceptions and mistakes

Inverse variance is not inverse standard deviation; confusing them changes the relative weights. Another mistake is assuming small variance means small total error when bias is present. Correlated estimates do not provide as much independent information as the simple formula assumes. A variance estimated from too little data can also make a weight unstable. Finally, the optimality statement concerns a specified class of estimators and assumptions, not universal superiority over every possible robust, nonlinear or task-specific combination method.

Business use

Precision weighting can support measurement fusion and repeated-estimate aggregation when the inputs are comparable. Teams should document the common target, calibration basis, variance estimates and any shared error sources. Validate uncertainty calculations as well as the combined point estimate. Monitor whether noise behavior changes with operating conditions, since fixed weights may become inappropriate. If one input dominates, investigate whether its apparent precision is credible. A weighted average can reduce random variability while leaving systematic error untouched, so calibration and model checking remain separate responsibilities.

Check your understanding

What were the normalized weights in the example?

Answer: 0.8 and 0.2.

Why was the combined estimate nearer 10 than 14?

Answer: Estimate A had the smaller variance.

What would undermine the independent combined-variance formula?

Answer: Shared or correlated errors, among other violations of its assumptions.

Related terms

Learn next

Repeat the two-estimate calculation with equal variances and observe the resulting equal weights. Then introduce a shared error term and compare the true combined variance with the independent formula. Keep bias separate from random variation throughout the exercise.

Variance

Sources

University of California, Berkeley, Statistics 210A, precision-weighted estimation exercise: https://stat210a.berkeley.edu/fall-2025/homework/homework11.pdf The two-measurement calculation, inventory-counter analogy and imaging-team case are original educational illustrations. The minimum-variance claim is explicitly restricted to independent unbiased estimates of a common quantity and linear unbiased combination.

Previous: Inverse transform sampling Next: Inverse-Wishart distribution