Is this the same as an ordinary random variable?

No. It represents uncertainty through paired possibility distributions in the specified measurement framework.

Does repeated measurement eliminate the internal component?

Not generally; repeating observations does not automatically resolve a residual systematic uncertainty.

Is an alpha-cut always a confidence interval?

No. Any confidence interpretation requires additional construction and assumptions.

AI Glossary / Graphs and symbolic AI / Advanced

Random-fuzzy variable

Definition

A random-fuzzy variable, in the measurement-uncertainty framework developed by Ferrero and colleagues, represents a measurement result while retaining separate information about random and nonrandom uncertainty contributions. It uses an internal possibility distribution for nonrandom contributions and an external possibility distribution for the combined uncertainty. This is a specific technical framework, not a generic synonym for any mixture of randomness and fuzziness.

What it means

Repeated measurements can vary because of random effects, while uncertainty about an imperfectly corrected instrument offset can remain across repetitions. Combining everything into one number may hide this distinction. A random-fuzzy representation preserves it so later calculations can treat contributions according to the selected uncertainty model.
Possibility distributions describe which values are compatible to different degrees with the available information. They are not ordinary probability densities, and a membership value should not automatically be read as a frequency or probability. Similarly named ideas, including fuzzy random variables and random fuzzy variables in credibility theory, use different definitions. The measurement-specific meaning is the one used here.

How it works

The analyst first identifies relevant measurement information and separates contributions under the chosen model. An internal distribution represents nonrandom uncertainty, including residual systematic contributions. A random distribution represents the random contribution in possibilistic form. Combining the internal and random distributions produces the external distribution.
One construction described by the framework uses a supremum of minimum-combined possibilities over potential shifts of the random distribution. At a selected membership level, interval boundaries provide a convenient way to inspect the result. Propagation and conditioning require the framework's stated operators and dependence assumptions; merely drawing two curves does not establish a valid uncertainty calculation.

Three practical examples

A measurement study can retain uncertainty about calibration offset separately from variability among repeated sensor readings.
A laboratory software prototype can show how the internal and external uncertainty intervals change after a transformation of a measured quantity.
An industrial research project can compare this representation with a probabilistic uncertainty analysis, checking how each expresses the same available evidence and assumptions.

Think of it this way

Think of a target whose center may be shifted within a band while individual shots also scatter around that center. The two features describe different sources of uncertainty. The analogy is limited because a possibility band is not automatically a probability distribution over the center, and the framework's mathematical combination must be specified rather than inferred from the picture.

Worked example

Consider an original illustrative measurement centered at 10 units. Model the internal distribution as rectangular: possibility 1 for center values from 9.8 to 10.2 and 0 outside. Model the random contribution around mode 10 as a triangular possibility distribution with peak 1 at 10, declining linearly to 0 at 9.8 and 10.2. These are chosen possibility shapes, not claims inferred from observed laboratory data.
At membership level alpha = 0.5, the internal interval is [9.8, 10.2]. The triangular random distribution has interval [9.9, 10.1], corresponding to deviations of −0.1 and +0.1 from its mode.
Under the shift-and-supremum/minimum construction, the external interval at this level extends the internal interval by those deviations. Its lower bound is 9.8 − 0.1 = 9.7 and its upper bound is 10.2 + 0.1 = 10.3. The four boundaries are therefore (9.7, 9.8, 10.2, 10.3): external lower, internal lower, internal upper, and external upper.
At alpha = 1, the random interval collapses to its mode and the external core matches the internal core [9.8, 10.2]. At alpha = 0.25, the random half-width is 0.15, giving external interval [9.65, 10.35]. The lower membership level admits a wider interval.
This arithmetic demonstrates the selected representation. It does not establish a 50% or 75% confidence interval without a justified probability-to-possibility construction and the relevant interpretation. The internal interval is also not removed merely by averaging more observations.

Illustrative case study

Fictional case: Laurel Measurement Lab compares repeated readings from a demonstration sensor with an uncertain residual offset. Its report previously displayed only one combined range, making it unclear which part might shrink with additional measurements.
The researchers prototype a random-fuzzy display showing internal and external distributions, document their selected shapes, and compare calculations with an independent uncertainty model. They retain the calibration assumptions in the report. This fictional case illustrates traceability of uncertainty sources; it does not claim that RFVs are required by a standard or universally more accurate than probabilistic methods.

Common misconceptions and mistakes

Do not interpret possibility 0.8 as an 80% probability without a specific justified connection. Alpha-cuts are defined by membership thresholds; they do not have a universal confidence interpretation. Do not replace the upper external boundary with a subtraction that makes the combined range narrower than its internal component in the worked construction.
Avoid equating nonrandom uncertainty with a known correction: a known correction and uncertainty about that correction are different. Also, random and nonrandom labels do not excuse ignoring dependencies. Applying an operator outside its assumptions can produce a tidy curve that misrepresents the measurement evidence.

Business use

The business relevance is mainly in specialist measurement, instrumentation, and uncertainty research. It may help teams preserve information about different uncertainty contributions when downstream decisions depend on how those contributions behave.
Before adopting the method, identify the measurement model, the evidence used to construct each distribution, and how results will be interpreted by decision makers. Compare outputs with an appropriate established analysis for the application. A general AI application should not add fuzzy terminology merely to make an uncertainty score appear more sophisticated; the representation must answer a concrete modeling need.

Check your understanding

Why is the upper external boundary 10.3 rather than 10.1 at alpha 0.5?

Answer: The positive random deviation of 0.1 extends the internal upper boundary 10.2 outward. Subtracting it would incorrectly contract that side of the range.

Related terms

Random error
Calibration
measurement uncertainty
possibility distribution
alpha-cut
fuzzy variable
systematic effect
theory of evidence

Learn next

Reduce the internal half-width from 0.2 to 0.05 while leaving the random distribution unchanged. At alpha 0.5, the internal interval becomes [9.95, 10.05] and the external interval [9.85, 10.15]. Explain which model component changed.

Sources

Sources: Ferrero, Prioli and Salicone, Conditional Random-Fuzzy Variables Representing Measurement Results, author manuscript: https://re.public.polimi.it/handle/11311/848749 . Ferrero and Salicone, foundational measurement RFV paper, institutional record: https://re.public.polimi.it/handle/11311/555634 .

Previous: Random forest Next: Random geometric graph