Este guia está disponível atualmente em inglês.
Technology / Technical guide
Parameter Identifiability: Unique Models & Hidden Ambiguity
Parameter identifiability asks whether different parameter choices can be distinguished through the model’s represented input and output. A model may reproduce observations while leaving its individual parameters ambiguous. More decimal places in an optimisation result do not resolve that ambiguity, because numerical precision and information content are different things.
An output can identify a combination without its parts.
Model relationships and the quality of represented information determine whether individual parameters can be distinguished.
01 / Model question
A parameter is an unknown part of a relationship.
Parameters can describe constants within a defined model, while states describe its evolving internal variables.
Parameters can describe constants within a defined model, while states describe its evolving internal variables. Identifiability concerns which parameter choices the represented observations can distinguish. The relationship between parameters, known inputs and outputs therefore comes before the fitting method.
The question is not merely whether there are more observations than parameters. Many observations can repeat the same dependency. Likewise, giving a coefficient a physical name does not establish that observations independently determine its value within the chosen model.
02 / Product ambiguity
Different pairs can produce exactly the same response.
Consider the hypothetical model y(t) = ab u(t), with a and b dimensionless unknown constants and input and output in the same units.
Consider the hypothetical model y(t) = ab u(t), with a and b dimensionless unknown constants and input and output in the same units. Parameters a = 1, b = 6 and a = 2, b = 3 both give y(t) = 6u(t).
Changing the input within this model cannot separate those pairs, because only their product appears. The combination ab can be determined while the individual factors remain ambiguous. A fitting program returning one pair has not created a distinguishing output relationship.
03 / Structural limit
Model structure can prevent uniqueness even ideally.
Raue and colleagues distinguish structural non-identifiability from practical limitations in their original paper’s abstract.
Raue and colleagues distinguish structural non-identifiability from practical limitations in their original paper’s abstract. Structural ambiguity involves parameter relationships within the model. It can remain with ideal, noise-free information because different choices are observationally equivalent.
In the product example, multiplying a by a nonzero factor and dividing b by that factor preserves the same response. Selecting a particular a by assumption can determine b algebraically, but that result uses the extra assumption rather than information newly supplied by y.
04 / Practical limit
Distinguishable models can still look nearly alike.
Practical difficulty concerns limited or uncertain information even when the ideal model permits a distinction.
Practical difficulty concerns limited or uncertain information even when the ideal model permits a distinction. Two parameter choices might produce slightly different outputs, but the difference can be small compared with represented measurement error or the range of available conditions.
This differs from exact equivalence. Additional information could matter in principle, although no particular data volume guarantees success. Close dependencies also make numerical estimates sensitive: a small change in the represented observations may produce a much larger change in individual fitted coefficients.
05 / State distinction
Observability assumes a different unknown quantity.
A state-observability question asks whether output history can distinguish possible internal states for a declared dynamic model.
A state-observability question asks whether output history can distinguish possible internal states for a declared dynamic model. Parameter identifiability asks whether that model’s parameter choices can be distinguished. Knowing one set of constants and inferring a state is not the same problem as inferring the constants.
A combined estimation problem can involve both, but naming it an observer does not establish uniqueness. Constraints, prior assumptions and fixed coefficients can narrow the permitted answer. Their contribution should remain visible instead of being mistaken for information independently contained in measurements.
06 / Scope
A unique numerical minimum need not remove model ambiguity.
An optimiser may select one answer through bounds, regularisation, starting values or a parameter convention.
An optimiser may select one answer through bounds, regularisation, starting values or a parameter convention. Those mathematical choices can make a numerical result reproducible while leaving a physical combination unresolved. The distinction is about explanation, rather than about the number of displayed digits.
No MARSTON parameter-identification function or unique dynamic model is claimed. This page gives no excitation, tuning or field-data procedures. It explains why identifiable combinations, practical sensitivity and state reconstruction must be considered as separate model properties.
Related technical guides: Least-Squares Fitting: Residuals, Weights & Input Errors · Model Residuals: Prediction Gaps & Ambiguous Causes · Phase-Locked Loops: Phase, Frequency & Synchronisation.
General background: Raue et al.: structural and practical identifiability — author abstract. These references explain industry concepts; they are separate from the applicable MARSTON product and project documents.
Key concepts and conditions.
Read each quantity with its reference, units and stated conditions. A general relation does not establish a model-specific value.
| Concept / quantity | Meaning / relation | Conditions / limits |
|---|---|---|
| Unknown parameter | Constant or coefficient within the model | Declared relationship and permitted parameter range |
| Equivalent output | Same response from different parameters | Compare the represented inputs and output model |
| Identifiable combination | For example, the product ab | Does not independently determine each factor |
| Practical sensitivity | Output difference relative to uncertainty | Requires a stated information and error model |
| State observability | Distinguishing internal initial states | Different unknown from a fixed parameter |
Engineering questions.
Do many observations guarantee identifiability?
No. They may repeatedly represent the same parameter combination.
Can ab be known while a and b remain unknown?
Yes. That is exactly the hypothetical product-model ambiguity.
Does a unique optimiser output prove uniqueness?
No. Bounds or assumptions may select one of several equivalent explanations.
Is structural ambiguity caused only by noise?
No. It can persist in an ideal noise-free model.
How is practical difficulty different?
The ideal distinction exists, but the available information may not resolve it reliably.
Is this the same as observability?
No. Parameter choices and dynamic states are different unknowns.
Technical enquiry
Discuss the technical question.
Parameter identifiability asks whether different parameter choices can be distinguished through the model’s represented input and output. Share the model and applicable operating or design information for an engineering discussion.