REVIEW 3 major objections 4 minor 3 references
Gravity inversion using a directional filtering method
T0 review · 3 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash
Pith's one-line read Directional filtering makes gravity inversion asymptotically exact
desk verdict Clever filter-based gravity inversion with appealing shallow-layer results, but the advertised asymptotic exactness is not established because the key update equation is admittedly wrong in the multi-cell case. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the directional filter vector $\mathbf{f}_{nm}=\mathbf{r}_{umn}/r_{nm}^b$, chosen with $b=2$, which makes the filter proportional to the forward coupling vector $\boldsymbol{\alpha}_{nm}$ between density cell $m$ and gravity point $n$. Its scalar product with the gravity correction vector isolates the component of the gravity-correction field along the line from the cell to the observation point, while the $r^{-b}$ decay gives more weight to nearby cells. This filter converts vector gravity equations into scalar update equations (Eq. (7)) in which every gravity datum contributes to every density correction, and it is the device that lets the method claim convergence of density only when gravity has converged.
What would settle it
Generate a synthetic model with a shallow anomaly and a deep anomaly that both contribute to the same gravity observation surface, run Eq. (7) on exact vector gravity until the gravity residual is at machine precision, and continue iterating; if the deep-cell density errors do not approach zero while the residual stays flat, the asymptotic-exactness claim is falsified.
Extended reading notes
Core claim
The central claim is that the density deviation field can be computed from presumed exact vector gravity deviations by iterating the under-relaxed scalar update $\delta\rho_m^{it+1} = \delta\rho_m^c + R\, \sum_n \mathbf{f}_{nm}\cdot(\delta\mathbf{g}_n^l-\delta\mathbf{g}_n^c) \big/ \sum_n (\boldsymbol{\alpha}_{nm}\cdot\mathbf{f}_{nm})$, with the gravity field updated afterward by the forward discrete Newton law. The derivation starts from Eq. (5), which is exact when only one density value is wrong; summing over all gravity points with a directional filter $\mathbf{f}_{nm}$ gives Eq. (6), which the paper acknowledges is usually false when several densities are wrong, and Eq. (7) then treats the resulting scalar as a correction. The claimed property is that $\delta\mathbf{g}_n^c=\delta\mathbf{g}_n^l$ is the only fixed point of the iteration in the relevant sense, so the method admits asymptotically exact solutions; the synthetic results show very accurate shallow densities and apparent, very slow convergence at depth.
Load-bearing premise
The load-bearing premise is that repeatedly applying the under-relaxed update (Eq. (7)) drives the density field to the true one even though the underlying correction equation (Eq. (6)) is derived from a single-error identity and is usually false when many density values are wrong; no convergence proof is given.
Editorial extensions
If this is right
- With exact vector gravity data and enough observation points, the iteration is intended to recover the density deviation field exactly, but only in the limit of infinitely many iterations.
- Shallow density anomalies can be recovered very accurately in practice, while deep density values converge at rates so low that practical runs leave tens of kg/m$^3$ errors.
- A gravity field reconstructed from the inverted density can be visually indistinguishable from the exact measured field even when deep densities are substantially wrong, so small gravity residuals do not by themselves certify a correct density model.
- The method as presented requires vector gravity data (or gravity-gradient data converted to vectors) and equal numbers of gravity and density unknowns, with interpolation and extrapolation suggested when data are scarce.
- The convergence rate depends on the filter-decay exponent $b$ and on the under-relaxation factor $R$; $b=2$ was used because it gives consistent convergence without further intervention.
Reading between the lines
- Editorial inference: A formal convergence proof would most naturally show that the update is a contraction in a norm where the forward map is coercive; the acknowledged inexactness of Eq. (6) for multiple density errors is the main obstacle.
- Editorial inference: A testable extension is to make the filter exponent $b$ and relaxation factor $R$ adaptive per depth or per iteration, since the paper identifies deep-cell convergence as the bottleneck and did not optimize $b$.
- Editorial inference: Because the method's failure mode is specifically deep-layer insensitivity, it could be combined with depth weighting or model regularization at depth, although the paper itself does not propose this.
- Editorial inference: The same directional-filter scalarization could be applied to gravity-gradient tensor components instead of vector gravity; the cited comparative studies of gradient-tensor components would provide a direct benchmark.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes an iterative gravity inversion method that uses the three Cartesian components of the gravity deviation vector. Starting from a discretized form of Newton's integral, the authors sum over all observation points and apply a directional filter f_nm to obtain a scalar update equation for each density cell. An under-relaxed iteration (Eq. 7) is then used to update the density deviation field, with gravity fields recomputed by the forward model at each step. The method is tested on two synthetic 100-cell models with exact gravity data. In both cases, shallow density deviations are recovered very accurately, while the deepest layer (z = -8750 m) retains errors of tens of kg/m^3 even though the reconstructed gravity field is visually indistinguishable from the exact one. The paper claims that the equations 'admit asymptotic exact solutions' and that deep densities 'apparently converge to the correct values at very low convergence rates.'
Significance. If the central claim were established, the method would offer a simple deterministic alternative to regularized potential-field inversion, with the notable feature of using full vector gravity data. The synthetic experiments are transparently described and the authors are honest about the method's limitations, including the lack of a convergence study. However, the asymptotic exactness claim is not supported: the derivation of the correction equation is explicitly admitted to be false when more than one density cell is simultaneously in error, and the two numerical examples do not demonstrate convergence to the exact deep density field, which is instead consistent with a null-space or alternative stable point. The paper would be a useful contribution if repositioned as a heuristic iterative scheme with documented behavior, but as written it overstates its theoretical grounding.
major comments (3)
- [Section 2, Eq. (6)-(7)] The derivation of Eq. (6) rests on Eq. (5), which is exact only when a single density deviation differs from its true value. The paper itself states that when density corrections at different locations are responsible for the gravity corrections, Eq. (6) 'ceases to be equivalent to Eq. (5) and both of them will usually be false.' Consequently, in the actual case of multiple erroneous cells, each update in Eq. (7) is contaminated by the other cells' errors. The statement that 'in the limit in which δg^c = δg^l, Eq. (7) accepts lim_{it→∞} δρ^{it+1} = δρ^l' only establishes that the true density field is a fixed point when the gravity residuals vanish; it does not establish that the iteration converges to that fixed point. No local stability or global convergence proof is supplied. This is a load-bearing gap for the abstract claim of asymptotic exact solutions, which therefore is not established by the manuscript.
- [Section 3, Cases 1 and 2 (Figs. 3, 4, 7, 8)] In both synthetic cases, density errors at the deepest layer remain large (tens of kg/m^3) while the reconstructed gravity errors are extraordinarily small. This behavior is equally consistent with a null-space or near-null-space of the forward operator as with slow convergence to the exact density. The text's conclusion that 'the evolution of the calculated results in the course of the iterative calculations suggests the possibility of exactness at the limit of an infinite number of iterations' is not supported by quantitative evidence. A convergence test is needed, for example a plot of the maximum or RMS density error versus iteration number on a log scale, showing monotone decay to machine precision for the deep layer, or a proof of unique invertibility of the discrete operator on the chosen grid.
- [Section 3, last paragraph before References] The convergence and uniqueness of the iteration are central to the paper's claims, yet the manuscript states that 'no study having been followed about convergence rates' and that the filter exponent b=2 was chosen because it 'apparently ensures consistent convergence.' Likewise, the under-relaxation factor R is said to vary automatically according to an unspecified error parameter, but the update rule is not defined. Without a sensitivity analysis over b and R, or at least a precise statement of the convergence conditions, the general applicability of the method beyond the two chosen synthetic cases remains unsubstantiated.
minor comments (4)
- [Section 3, Case 1] The sentence 'Eq. (7) is iteratively solved with R automatically varying according to one error parameter based on |δg^l - δg^c|' is too vague for reproducibility; please specify the exact formula used to adapt R.
- [Section 2, Eq. (8) discussion] The statement 'b=2, equivalent to use f_nm = α_nm' is inaccurate as written, because α_nm includes the factor -G δV_m in addition to the direction/distance dependence; the relationship is proportionality, not equality, and should be stated accordingly.
- [Throughout] There are several typographical and style issues: 'Fig.s' appears in Section 2, 'undistinguishable' in Section 3, and equation references are often given as 'Eq.s (7a)' or 'eqs.'; these should be standardized.
- [Section 3, Monte Carlo comparison] The comparison with Monte Carlo states only that 'observed overall convergence rates are similar, at least at the relatively small iteration numbers reached in MC calculations'; this is too vague to be informative and should either be removed or accompanied by quantitative results.
Circularity Check
No significant circularity: the synthetic tests use a standard forward-model self-check, and the fixed-point statement for Eq. (7) is a definitional property, not a fitted result presented as a prediction.
full rationale
The paper's derivation chain is not circular. Eq. (4) is a standard discrete forward model; Eq. (5) is exact only in the single-anomaly case, and the paper explicitly says that with multiple density errors Eq. (6) 'ceases to be equivalent to Eq. (5) and both of them will usually be false.' Eq. (7) is an under-relaxed iterative update whose fixed point, when the gravity residual vanishes, is the true density only because the true density is defined as the density that produced the exact gravity field. That is a trivial fixed-point identification, not a derivation of the target result from the target result. The synthetic tests generate gravity data from an arbitrarily chosen density field and then invert those data with the same forward model; this is standard synthetic validation, not circularity. The filter exponent b=2 is hand-selected and the relaxation factor R is adaptive, but these are method parameters, not fitted outputs renamed as predictions. There are no load-bearing self-citations: the references are to prior external work and no uniqueness theorem is imported from the authors' own previous papers. The paper's admitted lack of convergence analysis and its demonstration that very different deep density fields can produce nearly indistinguishable gravity fields are correctness/completeness limitations, not circular reasoning. Therefore the circularity score is 0.
Assumptions & free parameters
free parameters (2)
- b (filter exponent) =
2
- R (under-relaxation factor) =
adaptive, unspecified rule
assumptions (3)
- domain assumption The discretized forward equation (Eq. 4), with alpha_nm computed at grid centers, adequately represents the continuous Newtonian gravity integral.
- domain assumption Exact vector gravity data are available at the required points.
- domain assumption Matching the gravity field within V is sufficient to recover the density field in V.
Cite this review
Pith. "Pith review of Gravity inversion using a directional filtering method." pith.science (2026). https://pith.science/paper/QXXSBVXE
@misc{pith2026190807756,
author = {Pith},
title = {Pith review of: Gravity inversion using a directional filtering method},
year = {2026},
howpublished = {\url{https://pith.science/paper/QXXSBVXE}},
note = {Machine review of arXiv:1908.07756}
}
read the original abstract
The calculation of the underground density field from measured gravity data has been done by a variety of methods of varied types. The use of the vector gravity components is here addressed in order to develop one accurate gravity inversion method. The equation to directly calculate the Cartesian components of the gravity field from the density field is transformed in one non-exact correction equation. The discrete vector equations are transformed into scalar equations by using one vector function as directional filter. The used equations have the property that the convergence of the calculated density field to the exact values may only happen if the same occurs to the gravity field. The equations of the method so admit asymptotic exact solutions to the gravity inversion problem. The use of a smooth directional filter allows the equilibrated influence of all the gravity data over all the calculated density deviation corrections. The unknown density field is calculated by means of one iterative under-relaxed method. Two synthetic gravity fields are inverted by means of the presented method. The calculated shallow density fields are very accurate while the far density fields apparently converge to the correct values at very low convergence rates.
Reference graph
Works this paper leans on
-
[1]
Camacho, A.G., Fernández, J., and Gottsmann, J., 2011, The 3-D gravity inversion package GROWTH2.0 and its application to Tenerife Island, Spain: Comput Geosci 37(4), 621–633. Green, W.R., 1975, Inversion of gravity profiles by use of a Backus-Gilbert approach: Geophysics 40(5), 763–
work page 2011
-
[772]
Mosegaard, K., and Tarantola, A., 1995, Monte Carlo sampling of solutions to inverse problems: J
Li, Y., and Oldenburg, D.W., 1998, 3-D inversion of gravity data: Geophysics, 63(1), 109–119. Mosegaard, K., and Tarantola, A., 1995, Monte Carlo sampling of solutions to inverse problems: J. Geophys. Res. 100(B7), 12431–12447. Paoletti, V., Fedi, M., Italiano, F., Florio, G., and Ialongo, S., 2016, Inversion of gravity gradient tensor data: does it provi...
work page 1998
-
[2014]
Shamsipour, P., Marcotte, D., M
Evaluating the utility of gravity gradient tensor components: Geophysics, 79, G1–G14. Shamsipour, P., Marcotte, D., M. Chouteau, and Keating, P., 2010, 3D stochastic inversion of gravity data using cokriging and cosimulation: Geophysics, 75(1), 1–10. Zhdanov, M., Ellis, R., and Mukherjee, S., 2004, Three-dimensional regularized focusing inversion of gravi...
work page 2010
Reviewed August 14, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.