Does the method always recover point values?

No. Its estimates are interpretable through their averaging kernels, which may remain broad.

Can a sharper estimate be less certain?

Yes. Weights that isolate a location can amplify measurement noise.

Does normalization make every estimate unbiased?

No. It ensures the chosen normalization property, not exact recovery for every possible model.

AI Glossary / Mathematics and optimization / Advanced

Backus - Gilbert method

Definition

The Backus - Gilbert method is an approach to inverse problems that estimates local averages of an unknown quantity from indirect measurements. It chooses measurement weights so that the resulting averaging kernel is concentrated near a target location, while controlling sensitivity to measurement noise. The central question is not simply “What model fits the data?” but “What local information can these measurements resolve reliably?”

What it means

In an inverse problem, measurements combine information from different places or components. A sensor reading might reflect a weighted mixture of material properties across a region rather than one point. Recovering a sharp point value can amplify small measurement errors.

Backus - Gilbert analysis makes this limitation visible through resolution. The averaging kernel describes which parts of the true model contribute to the estimate. A broad kernel means the result blends a wide region. A narrow kernel can improve localization, but obtaining it may require weights that greatly magnify noise.

How it works

Start with a known linear relationship between the unknown model and the measurements. At a target location, form an estimate as a weighted sum of the data. Applying those same weights to the measurement sensitivity functions produces its averaging kernel.

Choose weights by minimizing a measure of kernel spread, with a normalization condition and a way to limit or penalize variance. The measurement covariance determines the estimate's noise variance. Repeat for other targets if a profile is needed.

Report both the resulting estimate and its resolution and uncertainty. Different spread definitions and trade-off settings lead to different estimators. The method cannot create resolving information absent from the measurement system.

Three practical examples

1. A geophysical study estimates a local average of an underground property from measurements with overlapping sensitivity regions, reporting how broadly each estimate is averaged.

2. A remote-sensing analysis examines whether two nearby layers can be distinguished or whether their contributions are inevitably blended.

3. An inverse-problem teaching exercise compares a sharply localized reconstruction with a more stable averaged estimate, making the noise penalty of sharpening explicit.

Think of it this way

Think of listening through two microphones that each capture both speakers in a room. Combining their recordings may emphasize one speaker, but aggressive subtraction can also magnify background noise. The analogy captures mixing and noise amplification. It does not mean that the method is automatically a general speech-separation system or that every inverse problem is linear.

Worked example

Consider two unknown values m1 and m2. Two measurements satisfy d1 = 0.6m1 + 0.4m2 + noise1 and d2 = 0.4m1 + 0.6m2 + noise2. Assume independent noise terms, each with variance one.

Estimate the first value using a d1 + b d2, with a + b = 1 so a constant model is reproduced. The averaging weights on the unknowns are r1 = 0.6a + 0.4b and r2 = 0.4a + 0.6b. They describe what the estimate actually averages.

Choosing a = 3 and b = −2 gives r = (1,0), perfect isolation of m1 in this model. But its noise variance is a squared plus b squared = 13. Choosing a = b = 0.5 gives r = (0.5,0.5) and variance 0.5. That estimate is much quieter but equally blends both unknowns.

For an intermediate example, minimize r2 squared + 0.02(a squared + b squared), subject to a + b = 1. Write a = 0.5 + 2.5t and b = 0.5 − 2.5t. The objective becomes 0.25(1−t) squared + 0.02(0.5 + 12.5t squared). Differentiating gives t = 0.5, hence a = 1.75 and b = −0.75.

The resulting kernel is (0.75,0.25), and variance is 3.625. If m1 = 10 and m2 = 2, noiseless data are 6.8 and 5.2; the estimate is 8. It is a local average, not an exact recovery of 10. This discrete teaching example uses a specified spread penalty; it is not a universal Backus - Gilbert formula.

Illustrative case study

Fictional case: Alder Survey receives a reconstructed subsurface profile with several sharp-looking peaks. Its reviewers ask for the averaging kernels and find that neighboring displayed points blend substantially overlapping regions.

The team changes the report to show the effective resolution alongside the estimates and explores a range of noise - resolution trade-offs. It also tests synthetic profiles whose true structure is known, keeping those tests separate from field observations. The case illustrates how an attractive plot can overstate detail if the inverse method's resolving limits are omitted; it reports no actual survey finding.

Common misconceptions and mistakes

Do not call a narrowly spaced display grid high resolution without checking the kernels. More plotted points do not necessarily represent more independent information.

Normalization does not eliminate all bias. It preserves specified reference behavior, such as a constant model, while a varying true model can still be averaged away from the target value.

Negative data weights are not automatically an error, but they can accompany substantial noise amplification. Also distinguish measurement uncertainty from uncertainty in the forward model: an incorrect sensitivity matrix can invalidate a precise-looking reconstruction.

Business use

The method is useful when decisions depend on what an indirect measurement can actually distinguish. For an engineering report, request the measurement model, covariance assumptions, trade-off choice, and resolution evidence with the reconstructed values.

Choose the desired locality in relation to the decision. A broad but stable regional average may answer one question well and be unsuitable for locating a small defect. Compare candidate trade-offs using synthetic tests and domain knowledge rather than selecting the sharpest-looking image.

Check your understanding

Why is the estimate 8 in the worked intermediate case when m1 is 10?

Answer: Its kernel is (0.75,0.25), so it averages 0.75 × 10 + 0.25 × 2 = 8. The result deliberately includes some contribution from m2.

Related terms

Inverse problem
Regularization
Covariance
averaging kernel
resolution matrix
linear estimator
ill-posed problem
noise amplification

Learn next

Study linear inverse problems next. Recalculate the example with correlated measurement errors and explain why variance is then a quadratic form involving the full covariance matrix, rather than just a squared plus b squared.

Covariance matrix

Sources

Primary research reference opened: Bill Menke, “Backus-Gilbert Resolution Computed from Ensembles,” https://www.ldeo.columbia.edu/users/menke/research_notes/menke_research_note196.pdf. It discusses localisation and resolution–variance trade-offs. The discrete objective, arithmetic and fictional case here are original hand-checked illustrations.

Previous: Backup validation Next: Backus-Naur form