Does the theory eliminate sampling error?

No. It helps identify, quantify where appropriate, and control its components.

Why does particle size matter?

A small number of compositionally important large particles can dominate a small portion's content.

Is the fundamental-error equation a complete sampling plan?

No. Lot boundaries, increment selection, handling, analytical methods, and decision requirements still need specification.

AI Glossary / Statistics and probability / Advanced

Gy's sampling theory

Definition

Gy's sampling theory is a framework for understanding and controlling errors when taking and preparing samples of heterogeneous particulate materials. Developed by Pierre Gy, it connects the physical properties of a material and the sampling procedure with the representativeness and uncertainty of the resulting measurement. It is especially relevant to ores, soils, powders, and other mixtures whose particles differ in size or composition. It is not a general theory of selecting training rows from a digital dataset.

What it means

A laboratory may measure a tiny portion of material accurately while that portion poorly represents the original lot. The difficulty arises before the analytical instrument does its work: valuable, contaminated, or otherwise important particles may be unevenly distributed or preferentially included.

Gy's framework makes the sampling chain part of the measurement problem. Collection, splitting, grinding, transfer, storage, and final subsampling can each affect the result. More precise laboratory analysis cannot reconstruct particles that were systematically excluded upstream.

The central practical question is whether the portion analyzed adequately represents the defined material for the intended measurement. That requires both an appropriate sampling design and physically correct execution.

How it works

The theory distinguishes error mechanisms such as fundamental sampling error, grouping and segregation, incorrect increment boundaries or extraction, and preparation losses or contamination. Fundamental sampling error arises from particle-level heterogeneity even under correct selection; other errors can reflect how particles are arranged or handled.

A commonly used approximation expresses the relative variance of fundamental error as v = K*(1/Ms − 1/ML), where Ms is sample mass, ML is lot mass, and K summarizes relevant material properties. In a more detailed formulation K includes composition, liberation, particle-shape and size-distribution factors, and a cubic particle-size term.

The factors, units, and assumptions must be checked for the material. The equation estimates one component of uncertainty, not total analytical error or bias. Increasing sample mass or reducing particle size can reduce that component, while appropriate increment collection and handling address other mechanisms.

Three practical examples

A mineral-processing team designs a sample-reduction sequence so a small assay portion can represent a much larger ore lot.
A soil laboratory evaluates whether its subsampling method preferentially selects fine particles and distorts contaminant measurements.
An industrial AI team reviews the physical sampling process behind its quality labels before blaming prediction errors entirely on the algorithm.

Think of it this way

Think of taking a spoonful from a jar containing large nuts and fine seasoning. A narrow scoop, a sample taken only from the surface, or seasoning lost during transfer can change the measured mixture. Weighing the spoonful to many decimal places does not fix those selection problems. The analogy explains heterogeneity; it is not a food-sampling protocol.

Worked example

For an original arithmetic illustration, assume a validated material-specific constant K = 4 grams and a lot mass ML = 10,000 grams. With a 100-gram sample, the approximate relative variance is 4*(1/100 − 1/10000) = 0.0396. The corresponding relative standard deviation is sqrt(0.0396) = 0.1990, or 19.90%.

Increasing the sample mass to 400 grams gives 4*(1/400 − 1/10000) = 0.0096, with relative standard deviation 9.80%. The standard deviation is roughly halved, not divided by four.

For a target relative standard deviation of 10%, solve 0.01 = 4*(1/Ms − 1/10000). This gives Ms = 1/0.0026 = 384.62 grams. Independent calculation confirms all results. The assumed K is invented for teaching; these masses are not recommendations for any real material, and the calculation excludes other sampling and analytical errors.

Illustrative case study

In a fictional recycled-materials company, composition measurements fluctuate between laboratory repeats. Managers initially propose a more expensive instrument. A sampling review instead finds that the small scoop used for subsampling cannot readily admit the coarsest fragments.

The laboratory redesigns the preparation and splitting procedure, retains more material at the relevant stage, and tests the revised process with suitable replicate samples. The AI team records which protocol produced each training label so that measurements from different procedures are not silently treated as equivalent. The company still evaluates instrument performance separately. This scenario illustrates a diagnostic sequence; it does not assert a quantified improvement or establish a universal equipment choice.

Common misconceptions and mistakes

Do not equate a large sample with an unbiased sample. A larger scoop from an unrepresentative location may remain systematically misleading. Mixing is not a guarantee of permanent homogeneity because particles can segregate again.

Fundamental-error variance is not the same as total measurement uncertainty. A formula calculated with guessed material factors can create false precision, and crushing can change material behavior or introduce contamination if poorly controlled. The theory's sampling terminology also differs from statistical sampling of independent people or digital records; transferring a mass formula into those settings is inappropriate.

Business use

For businesses whose decisions depend on physical assays, review the entire evidence chain from lot definition to reported value. Record sample masses, preparation stages, equipment, particle-size assumptions, and the basis for uncertainty estimates. Compare replicate results using a design that can distinguish collection, preparation, and analytical variation.

For machine-learning projects, this work improves understanding of target-label reliability. A highly accurate model trained against systematically biased assays may reproduce the measurement process rather than the material property management intended to predict.

Check your understanding

In the example, what happens when sample mass increases fourfold?

Answer: Relative variance falls from 0.0396 to 0.0096, while relative standard deviation falls from about 19.90% to 9.80%.

Does that prove total error is below 10%?

Answer: No; only the modeled fundamental-error component was calculated.

Related terms

representative sampling
fundamental sampling error
heterogeneity
segregation
composite sample
subsampling
measurement uncertainty
data provenance

Learn next

Next, map a real sampling and preparation chain. Identify where selection can become preferential, then study material-specific validation of fundamental-error models and replicate designs that separate major sources of variation.

Sources

Sources: U.S. EPA, “Correct Sampling Using the Theories of Pierre Gy,” https://health.hawaii.gov/heer/files/2021/07/USEPA1999c.pdf; U.S. EPA, “Guidance for Obtaining Representative Laboratory Analytical Subsamples from Particulate Laboratory Samples,” https://nepis.epa.gov/Exe/ZyPURL.cgi?Dockey=2000GTWM.TXT. Official guidance was opened and the error distinctions and mass-variance relationship checked. The numerical constant and business scenario are explicitly illustrative.

Previous: Guttman scale Next: H-spread