Non-Identifiability in Scientific Modeling

In the quest to understand the world, scientists and engineers build mathematical models to connect the reality we cannot see with the data we can measure. We rely on these models to reveal the mechanisms of disease, predict the behavior of complex systems, and reconstruct the history of life. However, a fundamental and often subtle challenge lurks within this process: ​​non-identifiability​​. This problem arises when our data, no matter how precise, are an ambiguous shadow of the underlying truth, allowing different—even contradictory—mechanistic stories to be equally valid. Ignoring this ambiguity can lead to flawed interpretations, incorrect conclusions, and a false sense of understanding.

This article provides a comprehensive guide to this critical concept. It is designed to equip researchers with the knowledge to recognize, diagnose, and address non-identifiability in their own work. The first section, ​​Principles and Mechanisms​​, will dissect the concept itself, distinguishing between fundamental structural flaws and practical data limitations. The second section, ​​Applications and Interdisciplinary Connections​​, will then illustrate how this single issue manifests across a diverse landscape of scientific fields—from pharmacology and epidemiology to engineering and evolutionary biology—and how researchers are developing clever strategies to overcome it.

Imagine you are a detective standing over a crime scene. The only clue is a perfectly circular shadow cast on the ground. What object cast it? It could be a sphere. But it could also be a flat disc, or a cylinder standing on its end. From this single piece of evidence—the shadow—the true shape of the object is ​​non-identifiable​​. You have run into a fundamental problem that plagues scientists and engineers across every discipline: sometimes, the data we can observe are an ambiguous projection of the reality we wish to understand. Different underlying truths can produce the exact same evidence.

This is the essence of ​​non-identifiability​​. It is not about having too much noise in our measurements or making a mistake in our calculations. It is a more profound issue, a kind of mathematical blind spot inherent in the way we model the world and the way we choose to observe it. Understanding this concept is not just a technical exercise; it is a lesson in scientific humility and a guide to designing more clever and insightful experiments. We will see that this single principle appears in guises as varied as drug processing in the human body, the spread of diseases, the evolution of life, and the regulation of our very genes.

Non-identifiability comes in two main flavors: structural and practical. It is crucial to distinguish them, for they are like the difference between a crime that is impossible to solve in principle and one that is merely difficult to solve with the current evidence.

The Perfect Crime: Structural Non-identifiability

​​Structural non-identifiability​​ is the "perfect crime." It is a fundamental property of the mathematical model itself, in combination with a chosen experimental setup. It means that even with perfect, noise-free, and continuous data, there exist multiple, distinct sets of model parameters that produce the exact same observable output. The problem is baked into the structure of our theory. No amount of data of the same kind, no matter how precise, can resolve the ambiguity.

Let's look at some of these "perfect crimes" in action.

​​1. The Inseparable Partners: Parameter Lumping​​

Consider a simple model of how a drug is processed in the body, a field known as pharmacokinetics. A drug is administered, enters a central compartment (like the bloodstream), and from there it can be eliminated or move to other tissues. A minimal model might look like this:

Here, $x_1(t)$ and $x_2(t)$ represent concentrations in two stages of a biological cascade, $u(t)$ is the drug input, and $y(t)$ is the final measured effect we observe. The parameters—$k_e, k_{tr}, k_d, \alpha$—are rates and scaling factors that describe the underlying biology. If we analyze how the output $y(t)$ depends on the input $u(t)$, we find that it is governed by a ​​transfer function​​ in the Laplace domain that looks like this:

$H(s) = \frac{Y(s)}{U(s)} = \frac{\alpha k_{tr}}{(s+k_e)(s+k_d)}$

Look closely at the numerator. The parameters $\alpha$ and $k_{tr}$ only appear as a product, $\alpha k_{tr}$. They are like two business partners who only ever contribute to a joint account. From the final balance, we can only ever know their total contribution, $\alpha k_{tr}$. We can never know how much each partner put in individually. A scenario where $\alpha=2$ and $k_{tr}=50$ produces the exact same output as one where $\alpha=100$ and $k_{tr}=1$. The parameters are "lumped" together and are structurally non-identifiable. Similarly, in the denominator, swapping the values of $k_e$ and $k_d$ leaves the expression unchanged. We can identify the two rates, but we cannot definitively assign them to their respective processes.

​​2. The Perfect Disguise: Symmetries​​

Another common source of structural non-identifiability is the presence of hidden symmetries in the model equations. Imagine a gene that regulates its own production. The protein, $p(t)$, represses the transcription of its own mRNA, $m(t)$, which in turn is translated to create more protein. A simple model for this negative feedback loop is:

Here, $\alpha$ is the maximum transcription rate and $\beta$ is the translation rate. Suppose our experiment can only measure the protein concentration, $p(t)$. It turns out there is a beautiful, hidden symmetry. For any positive constant $c$, if we consider a new set of parameters where the transcription rate is multiplied by $c$ (so $\alpha' = c\alpha$) and the translation rate is divided by $c$ (so $\beta' = \beta/c$), the model can produce the exact same protein output $p(t)$ by also scaling the hidden mRNA concentration.

A mechanism with very fast transcription ($\alpha$) and slow translation ($\beta$) is indistinguishable from one with slow transcription and fast translation. This is a perfect disguise. From observing only the final protein, we cannot unravel the individual contributions of transcription and translation. This is a critical failure, as it prevents us from understanding the true mechanistic strategy the cell is using. A similar problem famously plagues Age-Period-Cohort (APC) models in epidemiology, where the linear relationship $Age = Period - Cohort$ creates a symmetry that makes it impossible to uniquely separate the effects of aging, the current environment, and the generation you were born into.

​​3. The Unknowable Past: Functional Non-identifiability​​

Perhaps the most profound form of non-identifiability occurs in evolutionary biology. When we reconstruct the "tree of life" from the DNA of species living today, we are trying to infer a history of speciation and extinction events. We might model this with a birth-death process, where lineages branch (speciate) at a rate $\lambda(t)$ and terminate (go extinct) at a rate $\mu(t)$. The astonishing fact is that, based only on the reconstructed tree of survivors, there are infinitely many different historical scenarios—that is, different functions $\lambda(t)$ and $\mu(t)$—that could have produced the exact same tree.

An apparent "burst" of speciation early in a clade's history might be explained by a high speciation rate $\lambda(t)$ at that time. But it can also be explained by a model with a constant speciation rate and a declining extinction rate $\mu(t)$ over time. Both stories are equally consistent with the evidence of the survivors. The data we have—the extant phylogeny—only gives us information about a single composite function, a "pulled speciation rate," but not its two constituent parts. The true history is structurally non-identifiable.

The Case of Weak Evidence: Practical Non-identifiability

Structural non-identifiability is a flaw in the model or experimental concept. ​​Practical non-identifiability​​, on the other hand, is a flaw in the execution of an experiment. The model may be theoretically sound (structurally identifiable), but our real-world data—which is finite, noisy, and collected under specific conditions—may be insufficient to pin down the parameters with any reasonable precision. The crime is solvable in principle, but the evidence is too weak.

Imagine trying to determine the parameters of an enzyme's reaction speed, described by the famous ​​Michaelis-Menten​​ equation: $v = \frac{V_{\max} S}{K_M + S}$. The parameter $V_{\max}$ is the maximum reaction speed, and $K_M$ is related to the substrate concentration at which the speed is half-maximal. If we are lazy and only perform our experiment at very low substrate concentrations ($S \ll K_M$), the equation simplifies to a straight line: $v \approx \frac{V_{\max}}{K_M} S$. Our data will only allow us to identify the ratio $\frac{V_{\max}}{K_M}$, which is the initial slope. We cannot disentangle $V_{\max}$ and $K_M$ individually. We have made them practically non-identifiable by designing a poor experiment that doesn't probe the system's full range of behavior.

We can visualize this with a tool called the ​​profile likelihood​​. We fix one parameter (say, $V_{\max}$) at a range of values and for each value, we find the best possible fit for the other parameter ($K_M$). We then plot the quality of that best fit. For our myopic enzyme experiment, this profile would be nearly flat over a wide range of $V_{\max}$ values, indicating the data are indifferent to its value. In contrast, a well-designed experiment would yield a sharply peaked profile, zeroing in on a single best value. A broad but still curved peak indicates weak identifiability—we can estimate the parameter, but with large uncertainty.

This "weakness" of the evidence can be quantified using the ​​Fisher Information Matrix (FIM)​​. Intuitively, the FIM measures how much information our data provides about the parameters. It is constructed from the ​​sensitivities​​ of the model—how much the output $y(t)$ changes when we wiggle each parameter. If wiggling two different parameters, $\theta_1$ and $\theta_2$, produces nearly identical changes in the output, their sensitivities are highly correlated. The FIM becomes ​​ill-conditioned​​ (nearly singular), which is a mathematical way of saying our evidence is ambiguous. The inverse of the FIM approximates the uncertainty in our parameter estimates. An ill-conditioned FIM leads to a hugely elongated "uncertainty ellipse," meaning we can be very certain about one combination of parameters but almost completely uncertain about another.

Why does this mathematical subtlety matter so much? Because a model is more than just a curve-fitting tool; it is a vessel for our understanding.

The most critical consequence of non-identifiability is the ​​failure of mechanistic interpretation​​. In the gene regulation example, if we cannot distinguish between a "fast transcription, slow translation" strategy and a "slow transcription, fast translation" strategy, our model fails to tell us how the cell actually works. We have a black box that predicts correctly but explains nothing. Similarly, in control theory, the ability to observe a system's internal state can be destroyed if a parameter like a sensor's gain is unknown and non-identifiable.

Furthermore, non-identifiability can corrupt the process of ​​model selection​​. When comparing different models, we often use criteria like the ​​Akaike Information Criterion (AIC)​​, which balances goodness-of-fit against model complexity. A model with a non-identifiable parameter is needlessly complex; it has a knob that isn't connected to anything. The AIC correctly penalizes this "empty complexity," disfavoring a more complex model that offers no improvement in fit. Ignoring identifiability can lead us to choose overly complex and uninterpretable models.

Finally, non-identifiability can break the standard tools of ​​statistical inference​​. When testing hypotheses, such as whether a sample of neurons comes from one population or two, we often rely on theorems (like Wilks's theorem) that assume parameters are identifiable. In mixture models, this assumption is violated in a complex way, and the standard statistical tests (like the chi-squared test) give the wrong answer. Relying on them can lead to false discoveries. Special techniques, like the parametric bootstrap, are required to get a valid result.

Non-identifiability is not a death sentence for a modeling project. Rather, it is a call for rigor and careful thought. A systematic workflow can diagnose and often remedy these issues, turning ambiguity into clarity.

1. ​​First, Do the Theoretical Homework (Structural Analysis).​​ Before collecting a single data point, analyze the model equations. Use mathematical techniques (like transfer function analysis or differential algebra) to search for hidden symmetries, lumped parameters, or other sources of structural non-identifiability. If a problem is found here, the model must be fixed, either by ​​reparameterizing​​ into identifiable combinations or by planning a richer experiment (e.g., measuring an additional variable) that can break the symmetry.

2. ​​Second, Design a Clever Experiment (Practical Analysis and OED).​​ Once the model is structurally sound, the next step is to ensure it will be practically identifiable with real data. This involves ​​sensitivity analysis​​ and the ​​Fisher Information Matrix​​. The FIM should not be seen merely as a passive diagnostic tool, but as an active guide for ​​Optimal Experimental Design (OED)​​. We can use algorithms to design an input signal $u(t)$ and a sampling schedule that maximize the FIM's determinant or smallest eigenvalue, effectively designing an experiment that is maximally informative and pushes the system into regimes where the parameters' effects can be disentangled.

3. ​​Finally, Be Honest About Uncertainty.​​ If, due to practical constraints, some parameters remain weakly identifiable (their profile likelihoods are broad), the honest and scientific thing to do is to acknowledge this uncertainty. We report the large confidence intervals. We explore the range of possible mechanisms consistent with the data.

This journey from ambiguity to clarity shows that non-identifiability is not just a nuisance. It is a profound concept that forces a deep conversation between our theoretical models and our experimental reality. It pushes us to ask not just "What can we measure?" but "What do we need to measure to truly understand?" In answering that question, we become better scientists.

Having explored the mathematical heart of non-identifiability, we now embark on a journey to see where this "ghost in the machine" appears in the real world. We will find it lurking in hospital wards, in the chronicles of human history, within the microscopic machinery of viruses, and even in the grand enterprise of scientific discovery itself. Far from being an abstract nuisance, understanding non-identifiability is a key that unlocks deeper insights and inspires more ingenious science across a breathtaking range of disciplines. It teaches us not just about the limits of our knowledge, but about how to cleverly transcend them.

Imagine you are trying to understand a complex machine with a control panel full of knobs. You discover that two of the knobs are secretly linked; turning one also turns the other. You can observe the machine's output change, but can you ever say for sure how much of that change was due to the first knob versus the second? Of course not. You can only ever know their combined effect. This is the simplest and most common form of non-identifiability: perfect collinearity.

This exact situation arises with surprising frequency in statistical modeling. Consider epidemiologists trying to model the rate of bloodstream infections in hospital intensive care units. Their model might include a baseline infection rate (the "intercept," our first knob) and also a variable for whether the unit is accredited. Now, suppose that during the study period, every single unit is accredited. This "accreditation" variable is now constant, effectively becoming another baseline adjuster (our second knob). The model now has two knobs that are perfectly linked. Any attempt to estimate the unique contribution of the overall baseline rate versus the effect of accreditation is doomed to fail; only their sum is identifiable. The practical solution is simple: one of the redundant knobs must be removed, either by dropping the intercept or the constant accreditation variable.

While sometimes a result of a flawed model setup, this entanglement can also be a deep, unavoidable feature of the world. One of the most famous examples comes from epidemiology and sociology: the Age-Period-Cohort (APC) problem. Scientists want to disentangle three influences on a phenomenon like disease incidence or social attitudes:

1. ​​Age effect:​​ The consequences of biological aging or accumulating life experience.
2. ​​Period effect:​​ The impact of historical events that affect everyone alive at a particular time (e.g., a war, a pandemic, the invention of the internet).
3. ​​Cohort effect:​​ The influence of being born in a particular era, sharing formative experiences with your generation.

The trouble is, these three quantities are perfectly locked by a simple identity: $Period = Age + Cohort$. If you know a person's age and the current year, you know their birth year. Because of this linear dependency, you can't uniquely separate a linear trend in age from a linear trend in period and a linear trend in cohort. For example, is a rising disease rate in 60-year-olds over the last decade due to the aging process itself, something that happened in the last decade, or a characteristic of the generation born in the 1960s? The data alone cannot tell you. Any linear trend can be arbitrarily shifted between the three effects without changing the model's fit to the data one bit. This isn't a problem that can be solved with more data; it's a structural limitation. The solutions involve making explicit, scientifically-justified assumptions, such as constraining the trends of two effects to be zero, thereby attributing all linear change to the third. This is a profound lesson: sometimes, to get an answer, we must humbly admit what we cannot know from the data alone and instead impose a reasonable structure upon it.

Non-identifiability takes on a more subtle and elegant form in the world of dynamic systems—models describing how things change over time, often governed by differential equations. Here, the issue is often not simple collinearity but a hidden "scaling symmetry" in the equations themselves.

Let's venture into the world of mathematical virology. A classic model describes the battle between a virus and a host's immune system. Uninfected target cells ($T$) are infected by the virus ($V$), becoming infected cells ($I$). These infected cells then produce more virus at a certain rate ($p$), and eventually die. Scientists can typically only measure the viral load, $V(t)$, in a patient's bloodstream. The challenge is to infer the parameters of the underlying microscopic battle, such as the virus production rate $p$.

Herein lies a beautiful symmetry. Imagine a scenario where there are a certain number of infected cells, each producing virus at a rate $p$. Now, imagine a different, hypothetical scenario with twice as many infected cells, but where each cell produces virus at half the rate, $p/2$. What would the total viral load in the blood look like? It would be exactly the same! The observed dynamics of $V(t)$ are invariant to this scaling transformation. We can scale up the number of factories ($I$) and scale down the productivity of each factory ($p$) by the same factor, and the total output ($V$) remains unchanged. This means from blood measurements alone, we can never separately identify the number of infected cells and their individual production rate. We can only identify their product.

This kind of problem forces scientists to be more creative. The symmetry can be broken, but not with more of the same data. One must change the experiment. If we could somehow measure the number of infected cells $I(t)$ directly, or if we administered a drug that blocked viral production in a known way, we could break the symmetry and tell the two scenarios apart. This brings us to a crucial theme: ​​experimental design is one of our most powerful weapons against non-identifiability.​​

This principle is vividly illustrated in pharmacology, when determining a drug's effect. A drug's effect often lags behind its concentration in the blood, because it must travel to a target "effect site." This process is governed by an equilibration rate constant, $k_{e0}$. The drug's potency is described by its $EC_{50}$, the concentration that produces half of the maximal effect. At the very beginning of treatment, a drug with a fast equilibration rate ($k_{e0}$ is large) but low potency ($EC_{50}$ is high) can produce an effect that rises in exactly the same way as a drug with a slow equilibration rate ($k_{e0}$ is small) but high potency ($EC_{50}$ is low). The initial data can only identify the ratio $k_{e0}/EC_{50}$. How can we tell them apart? By designing a smarter experiment. Instead of a single dose, one could use a computer-controlled infusion to hold the drug's plasma concentration at several distinct plateaus. At each plateau, the system reaches a steady state, allowing for a clean measurement of the potency ($EC_{50}$). During the transitions between plateaus, the speed of the effect change is dictated purely by the equilibration rate ($k_{e0}$). By separating the experiment into parts that are sensitive to different parameters, we can measure each one individually.

In engineered systems, we often face the peculiar problem that our attempts to control a system can prevent us from truly understanding it. This is a central challenge in system identification, a field dedicated to building mathematical models of dynamic systems from experimental data.

Imagine you are tasked with creating a "digital twin" of a complex industrial robot, a model that can predict its every move. To do this, you collect data while the robot is operating. However, the robot is running under a feedback control system designed to keep it performing its task perfectly. The control computer constantly measures the robot's state (its position, velocity, etc.) and adjusts the motor commands to correct any deviation from the desired path.

The result is a dataset where the control input is always a deterministic function of the system's state. There's a perfect correlation between what the system is doing and the commands being sent to it. From this data, it's impossible to disentangle the robot's intrinsic dynamics (how it would move on its own) from the actions of the controller. You only observe the behavior of the combined, closed-loop system.

The solution, standard practice in control engineering, is wonderfully counter-intuitive: to learn about the system, you must stop controlling it so perfectly. You must intentionally "jiggle the steering wheel." By adding a small, random, but information-rich probing signal (known as a "persistently exciting" input) to the control commands, you break the deterministic feedback loop. This external signal provides the exogenous variation needed to excite the system's true dynamics, allowing you to see how it responds independently of the controller's influence and thereby build an accurate model of the robot itself.

Across these diverse fields, a unified set of strategies emerges for diagnosing and combating non-identifiability. A powerful diagnostic tool is the ​​sensitivity matrix​​. In computational mechanics, engineers build complex simulations (e.g., of sand grains in a landslide) with many micro-parameters, like inter-particle friction and stiffness. To calibrate these models, they construct a matrix that asks, for each parameter, "If I change this parameter by a small amount, how much do my macroscopic predictions (like the strength of the sand) change?"

This matrix can reveal two types of problems.

1. ​​Structural Non-Identifiability:​​ If a column of the matrix is all zeros, it means that changing the corresponding parameter has no effect on any of the measured outputs. For example, the coefficient of restitution, a parameter governing energy loss in dynamic collisions, has zero sensitivity in a quasi-static (very slow) test. It is fundamentally unidentifiable in that experimental context. The solution is to add a dynamic experiment, like a high-velocity impact test, which is sensitive to this parameter.
2. ​​Practical Non-Identifiability:​​ If two or more columns of the matrix are nearly parallel, it means their corresponding parameters have almost identical effects on the outputs. They are practically, if not structurally, unidentifiable. This high "collinearity" means that while a unique solution might exist in theory, in practice, any experimental noise will make it impossible to tell the parameters apart. The solution, again, is to add a new type of measurement that affects the parameters differently, breaking the parallel nature of their sensitivity vectors.

This brings us back to the power of ingenious experimental design. In semiconductor manufacturing, engineers use chemical additives called accelerators and suppressors to control the electro-deposition of copper wiring. In a simple experiment where the ratio of these chemicals is kept constant, their effects on the deposition current can become hopelessly confounded. To untangle them, engineers devise experiments that exploit other, differing physical properties of the two chemicals. For instance, if the molecules have different sizes, they will diffuse at different rates. Using a rotating disk electrode, which precisely controls diffusion, allows their effects to be separated. Alternatively, if they adsorb to the surface at different speeds, using time-resolved (chronoamperometry) or frequency-resolved (impedance spectroscopy) electrical measurements can distinguish the fast process from the slow one, and thereby attribute the effect to the correct chemical.

The consequences of non-identifiability can run even deeper, affecting the validity of the very tools we use for scientific inference. In the realm of machine learning, a Naive Bayes classifier might seem straightforward, but a hidden scaling non-identifiability lurks within. One can arbitrarily increase the "prior" belief in a certain diagnosis while simultaneously decreasing the "likelihood" score assigned to the evidence for it, and the final posterior probability can remain exactly the same. What resolves this ambiguity is a fundamental axiom of probability: all probabilities for all possible outcomes must sum to one. Enforcing this simple, bedrock constraint on the model's parameters eliminates the scaling freedom and makes the model identifiable.

Perhaps the most subtle implication comes from evolutionary biology, in the field of phylogenetics. Scientists build complex statistical models of DNA evolution to reconstruct the tree of life. They often use model selection criteria, like the Bayesian Information Criterion (BIC), to decide which model (e.g., a simpler one or a more complex one) is best supported by the data. BIC penalizes complexity, but its derivation rests on assumptions of model regularity—assumptions that are violated by non-identifiable models. For example, in a "mixture model" that allows different sites in a gene to evolve under different rules, the labels of the "rules" are interchangeable, leading to non-identifiability. When these assumptions fail, the standard BIC penalty becomes systematically too harsh. It over-penalizes complexity, creating a bias that favors simpler, but potentially incorrect, evolutionary models. This is a startling revelation: non-identifiability in our model of reality can mislead the statistical tools we use to judge it, potentially steering the course of scientific discovery in the wrong direction.

The journey through the world of non-identifiability leaves us with a sense of both humility and excitement. It is a reminder that data does not always speak for itself and that some questions, as posed, may be fundamentally unanswerable. Yet, it is this very challenge that pushes science forward. It forces us to think more deeply about the structure of our theories, to invent more powerful diagnostic tools, and to design more clever and insightful experiments. The ghost in the machine is not something to be feared; it is a teacher, a muse, and a constant invitation to be more ingenious.