There are many factors including the particle top size, the concentration and heterogeneity of the material in the stream. I suggest you have a look at the AS 4433.1-1997 as a starting point and a nice succinct example of what size? And what frequency! By experimentation from Dr Sans at http://is.gd/RaV8tq.

Ultimately it is about what is acceptable to the customer.To know how much you need to sample, you need to know how variable the stream you are sampling is and what is an acceptable level of uncertainty with regards to the results?

More often than not you will run into issues in both of these areas. Too few people understand the variability of their operations and reasonable expectation for the "acceptable level of uncertainty". Even if you have an expectation, the amount of sample typical for coal or iron ore (with the exception of fines products) is VERY large.

I would start with confirming the level of understanding that customers have of their uncertainty. Plant metallurgists and geologists often have a good handle on the potential range of grade (or ash) that occurs, sources of variability and conditions historically and in the future.

My own experience (coal and iron ore) with online analysis is that there is a disconnect between expectation for the units and the amount of sampling for appropriate calibration quality. It is this perception disconnect that must to be managed at as early a stage as possible (there are too many examples where it has been mismanaged!).

Thanks you've provided a nice example in DrSans' Variographic Experiment at Phosphate Hill. In our experience, the true variability of the material is not understood clearly, and consequently the sampling strategy does not provide results that are entirely representative in consideration of the compositional and size ranges of the material.

Can you recommend any papers that specifically illustrate the effects of different sampling increments on representativity? Many thanksyour comments about customer expectations are very apt. It is often the case that the operational results are seen as far more representative as they really are, due to a lack of definition of the true variability of the process stream being sampled.

All too often when we crunch the numbers we see the assumption that repeatability in the lab equates to representability of the bulk sample. Which to varying degrees is usually not the case?

When comparing teaspoons to tonnes it all comes back to having the compositional characteristics of the bulk stream clearly defined so that the sampling process ensures representativity. And you're right, all of this needs to be defined and understood to close that gap between expectations of operational performance and samplerepresentativity(and calibration sample requirements).

And the need for sample representivity continues post installation. Almost universally, where there is a representative sampling and analysis process, and a well calibrated analyser, there is good correlation and confidence in the data. A representative sampling process is pivotal.