C3 ยท Publication Volume 13

Element Associations and Ratios

pathfinders, multivariate patterns and the closure problem

Learning goals

The learner should be able to connect an element association to carriers and processes rather than correlation alone; recognise closure and dilution effects in compositional data; construct interpretable log-ratios with a geological numerator and denominator; handle zeros and censored parts honestly; distinguish redundant evidence from independent support; and validate multivariate patterns across batches, domains and new samples.

Elements co-vary because they share a source, mineral, fluid, transport pathway, adsorption surface, grain-size control, analytical interference, detection limit or closure constraint. Statistical association is therefore a hypothesis generator, not a process label. Interpretation must identify which mechanism predicts the sign, scale and domain of the association.

Process-based element associations

Begin with a carrier-and-reaction table. For each element, state likely mineral or phase, valence or speciation controls, mobility under relevant conditions, expected enrichment or depletion, and analytical recovery. Group elements by predicted process before inspecting correlations. Then test whether observed associations match, contradict or mix those groups.

A pathfinder is useful when its footprint is more detectable, broader or more stable than the target component and when alternative sources are controlled. It is not universally diagnostic. Ordinary lithology may contain the same element; weathering may decouple it; partial extraction may target a coating rather than bulk mineralisation. Use multiple lines of evidence and negative predictions.

Associations can change by domain and scale. Two elements may correlate within one lithology but not across lithologies, or show a global correlation driven only by different domain means. Report within-domain patterns and avoid inferring a shared process from pooled data alone.

Closure, dilution and log-ratios

When parts are expressed relative to a fixed total, increasing one part forces relative decreases in others even if their absolute masses do not change. This closure can create negative correlations. Variable dilution by an abundant component can make several trace elements rise and fall together without shared source enrichment.

Element associations become interpretable when carrier, process, closure, censoring and spatial domain are evaluated together
Element associations become interpretable when carrier, process, closure, censoring and spatial domain are evaluated together

Ratios compare relative abundance and can reduce a common dilution effect when the denominator is a stable reference for the process. For positive values, r=\ln(x_a/x_b) treats reciprocal changes symmetrically. Denominator choice is geological: a mobile or variably enriched denominator can create a false anomaly. Shared denominators also induce dependence among several ratios.

For a full positive composition x_1,\ldots,x_D, the centred log-ratio is

$\operatorname{clr}(x_i)=\ln\left(\frac{x_i}{g(x)}\right),\qquad g(x)=\left(\prod_{j=1}^{D}x_j\right)^{1/D}.$

The transformed parts sum to zero and remain dependent; balances or carefully selected ratios may be easier to interpret. A log-ratio cannot be calculated from a true zero, and censoring creates an interval rather than an exact transformed value.

Multivariate patterns and validation

Exploratory tools can reveal gradients, groups and outliers, but scaling, transformation, censoring treatment and missingness define the result. Standardising every element gives equal statistical weight to noisy or poorly measured variables. Retain a variable only when its quality and process meaning support the intended interpretation.

Cluster or component patterns should be tested against lithology, regolith, grain size, batch, method and spatial location. A component dominated by one batch boundary is a quality warning. A cluster identical to sampling medium is not necessarily a target class. Loadings or coefficients are model properties, not mineral-system facts.

Protect validation from spatial and temporal leakage. Fit transformations and pattern rules using training domains, then evaluate on separated samples collected or held out for that purpose. Report stability under reasonable censoring and scaling choices. A pattern that disappears under small defensible changes should not control a target decision.

Ratios, mass balance and mineralogical checks

A ratio can represent alteration mass change, mineral proportion, oxidation state proxy or source mixture only under explicit assumptions. Check conservation and immobile references where relevant. If both numerator and denominator are added or removed, ratio movement can be ambiguous. Use several ratios or a mass-balance model rather than one magic index.

Mineralogy tests the statistical story. If X and Y correlate because a carrier mineral contains both, grain-scale or phase observations should support co-location. If the association arises from oxide adsorption, a selective extraction and surface-phase indicator may strengthen it. If no plausible carrier exists, investigate units, detection limits, interference and batch processing.

Avoid double counting. Three highly correlated elements from one mineral may be one evidence group, not three independent votes. A target-ranking system should represent shared dependence and reward confirmation by genuinely different processes or media.

Worked synthetic example

Three synthetic samples contain X and Y values (20, 10), (40, 20) and (80, 40) mg/kg. The X/Y ratio is 2 in every sample, so \ln(X/Y)=\ln 2=0.693. Their absolute concentrations double, perhaps because the carrier amount increases, while relative composition stays constant.

Now add a dilution component so measured pairs become (10, 5), (40, 20) and (160, 80). The same ratio persists despite a sixteen-fold concentration range. If the decision concerns the X-bearing carrier relative to Y, the ratio is stable; if it concerns total X mass, the ratio hides important amplitude. Purpose determines the representation.

For a fourth sample, X is reported <2 and Y is 10 mg/kg. The ratio is in [0,0.2) under a non-negative bound, and the log-ratio has no finite lower bound as X approaches zero. Substituting X=1 would report \ln(0.1)=-2.303 with unjustified precision. The correct transformed result is censored or indeterminate until a suitable method resolves it.

Association audit workflow

  1. Define carriers, reactions and predicted associations before modelling.
  2. Verify units, methods, recoveries, limits and quality status by element.
  3. Separate geological, regolith, medium and analytical domains.
  4. Profile censoring and true or structural zeros before logarithms.
  5. Choose ratios or balances with a defensible process interpretation.
  6. Compare raw concentration, log-concentration and log-ratio views.
  7. Test patterns against grain size, lithology, batch and spatial structure.
  8. Group dependent elements so one process is not counted repeatedly.
  9. Validate transformations and pattern rules on separated observations.
  10. Seek mineralogical or multi-medium evidence for the proposed carrier.

Practice and review

  1. Construct a synthetic closed three-part composition where one constant absolute component appears negatively correlated with another after normalisation.
  2. Explain the geological assumptions behind choosing an immobile denominator.
  3. Derive interval bounds for X/Y when X<4 and Y=8 mg/kg.
  4. Design a check that distinguishes a carrier-mineral association from a batch-specific interference.
  5. Rewrite a ranking with four correlated elements as one evidence group plus an independent observation.

Review questions: What process or carrier links the elements? Could closure, dilution or limits create the pattern? Is the denominator stable? Are associations domain-specific? Has the pattern been validated independently?

Sources and further reading