Questions: Gravity Forward Modeling and Density Inversion

Question 1

A gravity survey shows a positive anomaly over a region. A geologist proposes a shallow, dense body and runs a forward model that matches the observed anomaly perfectly. She concludes the shallow dense body must exist. What is wrong with this conclusion?

• A. Forward modeling can only be done for spherical or cylindrical bodies, so irregular geometries cannot be tested
• B. Gravity inversion is non-unique — a deeper, larger, lower-density body could produce an identical surface anomaly
• C. A positive anomaly always indicates a shallow body; dense deeper bodies produce negative anomalies
• D. Perfect data fit means the model is verified; non-uniqueness only applies when the fit is imperfect

Answer: (B)

Non-uniqueness is a mathematical property of potential field inversion, not a data quality issue. Infinitely many density models produce the same surface gravity signal — a shallow, small, dense body and a deeper, larger, less dense body are indistinguishable from gravity data alone. A perfect fit to observations does not uniquely confirm a model; it only shows the model is consistent with the data. Eliminating alternatives requires additional constraints: seismic data, well logs, geological knowledge. This is why gravity interpretation always carries irreducible ambiguity without independent constraints.

Question 2

In gravity inversion expressed as d = Gm, what does the underdetermination of the system mean for the solution?

• A. The system has no solution because the number of equations exceeds the number of unknowns
• B. There is exactly one solution, but it requires very long computation to find
• C. There are infinitely many density models m that fit the data d equally well
• D. The sensitivity matrix G must be inverted, which is only possible if it is square

Answer: (C)

In gravity inversion, there are far more unknown cell densities (model parameters m) than gravity measurements (data d). The system has more unknowns than equations — it is underdetermined. The null space of G (density models that produce zero predicted gravity everywhere) is non-trivial, meaning you can add any null-space component to a fitting solution and still fit the data perfectly. Regularization selects one solution from this infinite family by imposing a preference (e.g., smallest total density contrast, or smoothest spatial variation). Without regularization, the inversion is ill-posed and returns no useful result.

Question 3

Given a sufficiently dense grid of surface gravity measurements, it is theoretically possible to uniquely determine the density distribution throughout the entire subsurface.

• T. True
• F. False

Answer: False

Non-uniqueness in gravity inversion is a fundamental mathematical property, not a sampling problem. Even with infinitely many perfectly precise surface measurements, the gravity field at the surface constrains only certain combinations of subsurface density — the projection of the density model onto the sensitivity matrix. The null space of that operator is infinite-dimensional: there are always infinitely many density distributions consistent with any set of surface observations. Adding more measurement points improves resolution and reduces some uncertainty, but cannot eliminate non-uniqueness. This contrasts with seismic travel-time tomography, which does approach uniqueness with sufficient coverage.

Question 4

Increasing the regularization parameter in Tikhonov regularization produces a smoother, simpler density model that may not fit the observed gravity anomalies as closely as a lower regularization parameter.

• T. True
• F. False

Answer: True

Tikhonov regularization minimizes a combined objective: data misfit + λ × model complexity. The parameter λ controls the tradeoff. High λ heavily penalizes model complexity (roughness or magnitude), pushing the solution toward smooth, featureless models that may not reproduce sharp anomalies in the data. Low λ prioritizes data fit, allowing geologically implausible spiky or oscillatory models. Neither extreme is useful: too little regularization overfits (noise is 'interpreted' as geology); too much underfits (real features are smoothed away). Choosing λ — via the L-curve, cross-validation, or geological judgment — is a central practical decision in applied geophysics.

Question 5

Explain why a density model that perfectly fits the observed gravity anomaly is not necessarily the correct representation of the subsurface, and how Tikhonov regularization addresses this problem.

Model answer: Perfect data fit is necessary but not sufficient for a correct model — this is the non-uniqueness problem. Because the sensitivity matrix G maps infinitely many density models to the same surface gravity, any particular fitting model has infinitely many equivalent alternatives. Without additional constraints, the mathematically 'best fit' model may be geologically meaningless (e.g., alternating high- and low-density cells that cancel to produce the right anomaly). Tikhonov regularization adds a penalty term that favors models with small density contrasts or smooth spatial variations, selecting the 'simplest' fitting model according to a chosen criterion. This does not resolve non-uniqueness — there are still infinitely many models — but it replaces an arbitrary choice with a principled one guided by geological plausibility. Independent geological or geophysical constraints (seismic reflectors, well data) are needed to further narrow the space of plausible models.

The philosophical principle is Occam's razor applied to geophysics: prefer the simplest model consistent with observations. Regularization operationalizes this preference mathematically. But choosing what 'simple' means (smooth? compact? minimum norm?) implicitly encodes geological assumptions that must be justified.