When buried glass, ash or cement weathers, some chemical elements leave the material and others are relatively enriched. Almost all existing attempts to "undo" this alteration treat it as a black-box regression from altered to pristine composition. This note starts from a simple observation. If composition is written in log-ratio coordinates (logarithms of each component relative to the average log of all components), pure leaching is approximately a straight-line shift. The set of shifts across many samples therefore spans a flat, low-dimensional set. Its dimension counts how many independent groups of elements were mobile. When alteration also adds material, for example secondary minerals or hydrothermal overprinting, the flat bends.
The intrinsic dimension of an alteration dataset, measured at the right observation scale, is therefore a physical quantity with a forward model behind it, not just a property of a point cloud. We propose to test whether scale-dependent dimension estimators can read this quantity from unlabeled, small geochemical populations. Fluid-inclusion paleotemperatures would serve as an independent physical check that the recovered direction tracks alteration intensity and not sample heterogeneity.
Core Claim
Let a pristine composition $x$ with $D$ components lose material so that component $i$ retains a fraction $f_i$. After the data are re-normalized to 100%, the centered log-ratio transform satisfies
$$\mathrm{clr}(x') = \mathrm{clr}(x) + \big(\log f - \overline{\log f}\,\mathbf{1}\big),$$
where $\overline{\log f}$ is the mean of the $\log f_i$ and $\mathbf{1}$ is the all-ones vector. The re-normalization constant drops out. If retention decays as $f_i = e^{-k_i t}$ with fixed relative rates $k_i$, the displacement is $-t(k - \bar k\mathbf{1})$, a line parameterized by exposure $t$. Several independently varying rate groups give a flat of that many dimensions. Adding a secondary phase is a convex mixture in the simplex, which is curved in these coordinates.
This yields three testable predictions. A population of altered samples should show a mesoscale dimension plateau equal to the number of independent leaching pathways. Matrices that drift in opposite directions should appear as a union of distinct flats, not one global subspace. Positive curvature on the plateau should signal addition processes, not removal. The opposite drift is documented: silica rises in weathered high-potassium glass and falls in lead-barium glass (Composition analysis and identification model of ancient glass products based on Spearman correlation coefficient and BP neural network).
What is new is not applying a dimension estimator to geochemistry. It is giving the estimate a physical ground truth that can be simulated, and checking it against an independent thermometer.
Why This, and Not the Obvious Version
The obvious combination is to run Gride or a diffusion-based estimator on the 14-oxide glass data and report a curve. The research file shows nobody has done this. Done naively it would be uninterpretable, for three reasons.
First, the populations are tiny, with 67 to 100 samples (Analysis of the identification results of the ancient glass products based on the progressive regression model). Second, the centered log-ratio space of 14 parts has rank 13, not 14. Third, below-detection zeros, such as barium in potassium glass, force an imputation that creates artificial lattice structure at fine scales.
Diffusion-model estimators such as FLIPD (A Geometric View of Data Complexity: Efficient Local Intrinsic Dimension Estimation with Diffusion Models) need a trained generative model. Their absolute estimates depend on network architecture, so they are not credible at this sample size, and we drop them. The usable estimators are distance-only ones. Gride (The generalized ratios intrinsic dimension estimator) probes coarser scales by using more distant neighbor pairs instead of subsampling the data, and it is implemented in DADApy (DADApy: Distance-based Analysis of DAta-manifolds in Python). The other is the volume-growth test used to detect cusps and branching (Token Embeddings Violate the Manifold Hypothesis).
Even these may not be identifiable at roughly 70 points. Measuring where they become identifiable is part of the contribution, not an afterthought.
Research Program
We would build a forward simulator of alteration in the simplex. It would cover pristine end-member populations with realistic within-type scatter and leaching with one to three independent rate groups. It would also include mixing-in of secondary phases (phosphate, iron, clay-like components) and polyphase histories in which leaching is followed by overprinting. Analytical noise and detection-limit zeros would be added.
Ground-truth dimension at each scale is known by construction:
- At the finest scales, isotropic noise dominates and the estimate approaches the full 13 dimensions.
- At intermediate scales, the estimate equals the pathway count.
- At the coarsest scales, matrix-type clusters collapse toward zero dimensions.
This mirrors the fine-scale noise inflation and coarse-scale recovery Gride shows on a noisy spiral, and the microband-versus-shearband separation obtained by tuning neighbor counts in On local intrinsic dimensionality of deformation in complex materials.
We then estimate, per sample and per scale, local dimension and local principal direction. From these we build a restoration operator: project a weathered sample back along its estimated local flat to the pristine-population boundary.
The physical anchor comes from fluid inclusions. In diagenetic systems such as the tuffaceous sandstones of Diagenetic Evolution Mechanism of the Jurassic Tuffaceous Sandstone Reservoir in Qikou Sag, Bohai Bay Basin, East China, homogenization temperatures in authigenic quartz bracket the temperature at which ash dissolved. That study uses those temperatures only as a qualitative threshold. We would instead pair SEM-EDS compositions of altered ash with homogenization temperatures of co-genetic inclusions measured in the same cement zones, with zones identified by cathodoluminescence imaging. We would then test whether a sample's position along the recovered alteration flat is rank-correlated with temperature, while its residual off the flat is not.
Inclusions can themselves be reset after trapping, because stretching raises homogenization temperature without changing ice-melting salinity. We therefore use the lowest homogenization temperature within each petrographically defined assemblage and discard assemblages whose ice-melting temperatures disagree.
No joint composition-plus-microthermometry dataset exists in the file, so this arm requires new measurements on roughly 10 to 20 thin sections. The Travis Peak work (A 48 m.y. history of fracture opening, temperature, and fluid pressure: Cretaceous Travis Peak Formation, East Texas basin) shows that trapping conditions from cemented fractures can be reconstructed to within 15% of reservoir pressure. That supports the thermometer as an independent axis.
Baselines
The naive reference point assumes no alteration structure: restore every weathered sample to the mean unweathered composition of its matrix type, and count pathways as zero. Any method that cannot beat class-mean restoration in Aitchison distance has learned nothing about alteration. Aitchison distance is the Euclidean distance between centered log-ratio vectors.
The strongest restoration competitor is the 14-input, 14-output back-propagation network of Composition analysis and identification model of ancient glass products based on Spearman correlation coefficient and BP neural network. It reports an "overall prediction error of less than 0.1 across more than 100 samples." The file does not say what that error is measured in, so we will reimplement it and not compare against the quoted figure. We will run it on identical leave-one-artifact-out splits in centered log-ratio coordinates, with the same hyperparameter search budget as our method (about 50 configurations). Beating it means lower held-out Aitchison error with non-overlapping bootstrap intervals, together with a stated pathway count that the network cannot provide.
For classification and subspace finding, the strongest existing methods are the following. Logistic regression with Fisher discriminant analysis, which classified all 67 known samples correctly (Analysis of the identification results of the ancient glass products based on the progressive regression model). Log-ratio transform plus 4-component PCA with density clustering, reporting 90% SVM accuracy, 91% logistic-regression accuracy and SVM F1 of 0.95 on 69 artifacts (Analysis and identification of the composition of ancient glass objects based on an improved DBSCAN model). Classification accuracy is already near ceiling, so we do not claim to beat it. Our question is pathway recovery.
The most dangerous baseline for pathway counting is PCA on weathered-minus-unweathered displacement vectors, fit separately per matrix type. Under the translation model it is the direct estimator. No reported result was found on Lacuna for pathway counting on any geochemical dataset.
Two ablations isolate the key ingredient. In the first, the identical pipeline runs on raw percentages. Re-normalization curves leaching paths in raw coordinates, so dimension should be inflated and spurious curvature appear. In the second, the translation constraint is removed from restoration and replaced with an unconstrained local regressor.
Why not just…?
Why not just run PCA on the log-ratio data, as the DBSCAN study does? Global PCA fits one subspace to all samples. Opposite drifts in the two glass types then become a single averaged direction, and the known branching is lost. The synthetic benchmark measures this directly: angle error of the recovered directions when two flats with opposite silica components are pooled. Per-type PCA on displacement vectors is a genuine threat, though. Whenever weathered and unweathered spots can be paired and types are known, it may match the scale-dependent estimators exactly for pure leaching. If the benchmark shows this, we will say so and confine the claim to the cases where pairing or type labels are unavailable, or where alteration is curved by addition.
Why not just train a better neural network to map weathered to pristine compositions? A network can minimize error without saying how many processes acted or whether material was removed or added. With 67 to 100 samples, overfitting is the acknowledged risk. The translation-constrained restorer has at most a few direction parameters per matrix type. The decisive comparison is held-out error at matched tuning budget. If the network wins, the geometry still supplies the count and the addition-versus-removal diagnosis, but restoration accuracy would not be part of our claim.
Why not just use bivariate cross-plots of temperature against composition? Cross-plots work when one variable drives alteration. But Testing the basin-centered gas accumulation model using fluid inclusion observations: Southern Piceance Basin, Colorado could not resolve systematic pressure trends from 55 to 110 MPa in such plots. Choosing which oxide pair to plot also presupposes the pathway. Our temperature test projects onto a data-derived direction. We will report the best single-oxide-ratio correlation with temperature alongside it, and if that matches, the geometric step adds nothing for this question.
Why not just fit a Gaussian mixture with full covariances? Elongated mixture components can approximate flats, and this is a fair competitor for pathway counting. Mixtures cannot represent curvature from addition processes, and they need the component count chosen in advance. We include a mixture model with component number selected by the Bayesian information criterion in every synthetic and real comparison.
Proposed Experiments
The first experiment is the synthetic identifiability study. We sweep sample size from 50 to 5,000, the number of pathways from 1 to 3, the noise level, the zero fraction, and the strength of addition. Metrics are:
- exact-match accuracy of the mesoscale plateau against the true pathway count;
- angle between estimated and true displacement directions;
- correct detection of curvature;
- restoration error in Aitchison distance.
Gride with scale set by neighbor rank, the volume-growth test, displacement-vector PCA, global PCA and the mixture model all run on identical draws. The decisive output is the smallest sample size at which each method recovers the pathway count 90% of the time. If no method reaches this at 70 samples, the glass-data analysis becomes a demonstration, not evidence.
The second experiment applies the pipeline to the archaeological glass dataset used by all three glass papers. It is pre-registered with one prediction: about one-dimensional flats per matrix type, with silica loadings of opposite sign. Restoration is scored by leave-one-artifact-out error against the back-propagation network.
The third experiment is the paired cement study. Here the decisive check is that position along the recovered flat correlates with minimum assemblage homogenization temperature (Spearman, with a permutation test), while the off-flat residual does not. A null on that check would mean the mesoscale direction reflects provenance heterogeneity and not alteration. We would report that outcome as a negative result for the physical interpretation.
References
- Composition analysis and identification model of ancient glass products based on Spearman correlation coefficient and BP neural network
- Analysis of the identification results of the ancient glass products based on the progressive regression model
- A Geometric View of Data Complexity: Efficient Local Intrinsic Dimension Estimation with Diffusion Models
- The generalized ratios intrinsic dimension estimator
- DADApy: Distance-based analysis of data-manifolds in Python
- Token Embeddings Violate the Manifold Hypothesis
- On local intrinsic dimensionality of deformation in complex materials
- Diagenetic Evolution Mechanism of the Jurassic Tuffaceous Sandstone Reservoir in Qikou Sag, Bohai Bay Basin, East China
- A 48 m.y. history of fracture opening, temperature, and fluid pressure: Cretaceous Travis Peak Formation, East Texas basin
- Analysis and identification of the composition of ancient glass objects based on an improved DBSCAN model
- Testing the basin-centered gas accumulation model using fluid inclusion observations: Southern Piceance Basin, Colorado