Non-Identifiable Models

In scientific inquiry, we often build mathematical models to understand the hidden mechanisms of the world, much like a detective trying to deduce the inner workings of a black box. We observe inputs and outputs to infer the machinery inside. But what happens when two entirely different internal designs—two different models or sets of parameters—produce the exact same observable behavior? This perplexing challenge is known as non-identifiability, a fundamental issue that forces us to confront the limits of what we can know from data. It addresses the critical knowledge gap that exists when our data are fundamentally ambiguous, leaving us with multiple plausible explanations for the same phenomenon.

This article provides a comprehensive exploration of non-identifiability. In the "Principles and Mechanisms" chapter, we will dissect the two primary forms of this ambiguity—structural and practical non-identifiability—using clear examples and visual aids like the profile likelihood to understand their origins and diagnostics. Subsequently, the "Applications and Interdisciplinary Connections" chapter will demonstrate how this concept is not a mere theoretical curiosity but a practical challenge and a source of insight across diverse fields, from systems biology and evolutionary studies to engineering and the regulation of new technologies. By understanding non-identifiability, we learn not only to be better modelers but also more insightful scientists, capable of turning ambiguity into a compass for discovery.

Figure 1: The shape of the profile likelihood reveals the identifiability of a parameter. A sharp peak (left) indicates a well-identified parameter. A shallow, wide curve (center) indicates practical non-identifiability—the data are weakly informative. A perfectly flat profile (right) indicates structural non-identifiability—the data are completely uninformative over a range of values.

Imagine you are a detective presented with a peculiar black box. You can put things in (inputs) and observe what comes out (outputs), but you cannot open the box. Your task is to deduce the machinery inside. You run some experiments, and you develop a theory—a beautiful model of gears and levers that perfectly explains all your observations. You are about to declare the case solved when your colleague walks in. She has a completely different model, based on springs and pulleys, that also explains every single observation, perfectly. Which model is correct? How can two different mechanisms produce the exact same behavior? This is not just a philosophical riddle; it is a profound and common challenge at the heart of modern science, known as ​​non-identifiability​​.

In science, our models are our theories of the machinery inside the black box of nature. The parameters of the model—the rate constants, the interaction strengths, the degradation rates—are the specifications of its gears and levers. When we fit a model to data, we are trying to deduce these parameters. But what if the data we can collect is fundamentally ambiguous? What if different sets of parameters, or even entirely different models, leave the exact same fingerprints?

This ambiguity isn't a single problem, but a spectrum. At one end, we have a deep, fundamental ambiguity, and at the other, a more practical, data-limited one. To truly understand how we build knowledge, we must learn to recognize and interpret these two kinds of shadows.

Let’s start with a wonderfully crisp example from the world of time series analysis. Imagine we are observing a noisy signal over time, and we find that its correlation from one moment to the next, its ​​autocorrelation function (ACF)​​, has a specific shape. We might propose a simple "moving average" model to explain this, of the form $X_t = W_t + \theta W_{t-1}$, where $W_t$ is random noise and $\theta$ is our key parameter. A little bit of mathematics shows that the crucial part of the ACF is determined by the term $\frac{\theta}{1+\theta^2}$.

Now, here is the twist. What if we test the value $\theta=4$? The ACF value is $\frac{4}{1+16} = \frac{4}{17}$. What if we test $\theta = 0.25$? The value is $\frac{0.25}{1+0.0625} = \frac{0.25}{1.0625} = \frac{4}{17}$. They are identical! In fact, for any value of $\theta$, the parameter $1/\theta$ gives the exact same autocorrelation function. Based on observing the ACF alone, we can never distinguish between $\theta$ and $1/\theta$. This is not a problem of noisy data or not having enough of it. It is a fundamental property of the model and what we chose to observe. This is ​​structural non-identifiability​​.

Now consider a different scenario. We are studying a chemical reaction, and we know that two rate constants, $k_{\text{fast}}$ and $k_{\text{slow}}$, govern the process. Our theory tells us that if we could measure the reaction with perfect precision, we could uniquely determine both rates. They are structurally identifiable. However, our instruments are noisy. Suppose the two rates are very close to each other, say $k_{\text{fast}} = 10.1$ and $k_{\text{slow}} = 10.0$. The subtle signal produced by their tiny difference is completely swamped by the random jitter of our measurements. We can find a "best" estimate, but the range of other plausible values is enormous. While we might be certain the rate is "around 10", we have no real confidence in the tiny difference between the two parameters. This is ​​practical non-identifiability​​.

The difference is like that between a perfect crime and a blurry photograph.

• ​​Structural non-identifiability​​ is the perfect crime. The perpetrator left no unique evidence. No matter how closely the detective looks, the clues point equally to multiple suspects. The problem is in the setup of the crime itself; more investigation of the same kind won't help.
• ​​Practical non-identifiability​​ is the blurry photograph. A single suspect is in the photo, but the image is so fuzzy that we can't be sure of their identity. The problem is the quality of the data. A sharper camera or a better-lit photo could solve the case.

Structural non-identifiability arises when the experimental design itself creates blind spots. It’s like trying to judge a 3D sculpture by looking at only its shadow. Different sculptures can cast the same shadow. This can happen in several ways.

​​1. Hidden Players in a Cascade:​​ Imagine a chain of command: a General ($u$) gives an order to a Colonel ($x_1$), who relays it to a Major ($x_2$), who finally orders a Captain ($x_3$) into action. We can see the General give the order, and we can see the Captain's final action. But let's say the Major's office is a black box; we can't observe the Major at all.

Our model might look like this simple cascade:

We can measure $x_1$ and $x_3$, but not $x_2$. From the data on $x_1$, we can figure out the Colonel's parameters, $k_1$ and $d_1$. But what about the others? It turns out that the link between the Colonel's message ($x_1$) and the Captain's action ($x_3$) depends only on the product of the rates, $k_2 k_3$, and the sum and product of the degradation rates, $d_2+d_3$ and $d_2 d_3$.

You can't determine $k_2$ and $k_3$ individually. If the real values were $(k_2=2, k_3=10)$, the product is $20$. A model with $(k_2=4, k_3=5)$ would give the exact same output for $x_3$, because the product is also $20$. Similarly, you can't tell whether the Major's decay rate is $d_2$ and the Captain's is $d_3$, or vice-versa. You can only identify the pair $\{d_2, d_3\}$. The unobserved intermediate, the Major, has confounded our ability to assign individual properties to him and the Captain. This is a common problem in complex biological networks, where not seeing a crucial intermediate molecule makes it impossible to disentangle the rates of the reactions surrounding it.

​​2. Insufficient Probing:​​ Sometimes, the problem isn't a hidden player but a limited perspective. Consider a simple process modeled by an exponential decay, $y(t) = \theta_1 \exp(-\theta_2 t)$. We want to find the initial amount $\theta_1$ and the decay rate $\theta_2$. What if our experiment consists of taking a single, perfect snapshot at time $t = 1$? We measure a value, say $y(1)=5$. Our equation becomes $5 = \theta_1 \exp(-\theta_2)$. This is one equation with two unknowns. An infinite number of pairs $(\theta_1, \theta_2)$ satisfy this condition (e.g., $\theta_2=1, \theta_1 \approx 13.6$ works, but so does $\theta_2=2, \theta_1 \approx 37$). We have a curve of possible solutions, not a unique point.

To solve this, we need more information. If we take another snapshot at $t=2$, we get a second equation, which allows us to solve for both parameters uniquely. Even better, if we film the entire process (i.e., measure the whole trajectory $y(t)$), we have more than enough information. This is why in enzyme kinetics, measuring the initial rate of reaction at just one substrate concentration is not enough to determine the two famous Michaelis-Menten parameters, $V_{\max}$ and $K_M$. You create a structural non-identifiability. To identify them, you need to measure rates at several different substrate concentrations or measure the full time course of the reaction.

​​3. Inherent Symmetries and Redundancies:​​ Sometimes we create non-identifiability ourselves by how we process the data. In our exponential decay model $y(t) = \theta_1 \exp(-\theta_2 t)$, what if we are only interested in the shape of the decay, not its absolute value? We might normalize the data by dividing by the initial value: $z(t) = y(t)/y(0) = \frac{\theta_1 \exp(-\theta_2 t)}{\theta_1} = \exp(-\theta_2 t)$. Now our observation $z(t)$ only depends on $\theta_2$! We have completely erased any information about $\theta_1$. We created a structural non-identifiability for $\theta_1$ by our choice of analysis. In more complex models, these redundancies can be deeply hidden. For instance, if a biological process is activated when a molecule $D$ exceeds a threshold $K$, the response might only depend on the ratio $D/K$. In this case, doubling the production of $D$ and simultaneously doubling the threshold $K$ might have no effect on the final output, making the absolute values of these parameters impossible to identify from that output alone.

How do we diagnose these problems in practice? One of the most powerful tools is the ​​profile likelihood​​. Imagine you are trying to find the best-fit value for a single parameter, say $k_T$, from a complex model. The "likelihood" is a measure of how well the model fits the data for a given set of parameter values. A higher likelihood means a better fit.

To create the profile likelihood for $k_T$, we march along a range of possible values for it. At each fixed value of $k_T$, we adjust all other parameters in the model to find the best possible fit we can achieve. We then plot this best-possible-likelihood as a function of the fixed $k_T$. The resulting curve is the profile likelihood, and its shape is incredibly revealing.

• For a ​​well-identified parameter​​, the profile will look like a sharp mountain peak. There is one clear value of the parameter that gives the best fit, and the likelihood drops off steeply as you move away from it.

• For a ​​practically non-identifiable parameter​​, the profile is a vast, shallow valley. There is a "best" value at the bottom, but the landscape is so flat that you can wander for miles in either direction without much change in altitude (likelihood). This tells you the data have a very weak preference for the "best" value; a wide range of values are almost equally plausible. This is the signature of the blurry photograph.

• For a ​​structurally non-identifiable parameter​​, the profile is perfectly flat over a range of values. It’s a high plateau, not a valley. Any value on this plateau gives the exact same, best-possible likelihood. This is the signature of the perfect crime. The data are completely indifferent to the parameter's value within that range, because some other parameters have adjusted to compensate perfectly.

Now that we have grappled with the principles of non-identifiability, you might be tempted to see it as a rather frustrating aspect of mathematical modeling—a pesky limitation that gets in the way of finding the "truth." But that would be missing the point entirely! In science, when Nature gives you a puzzle, it is also giving you a clue. Non-identifiability is not a roadblock; it is a signpost. It is the universe’s way of telling you, with perfect mathematical clarity, "Your question is ambiguous," or "Your experiment is not yet clever enough." It forces us to be more precise in our thinking, more creative in our experimental designs, and more honest about the limits of our knowledge.

Think of yourself as a detective. You have a collection of clues—the data—and a theory of the crime—the model. A non-identifiable model is like having a story where multiple suspects fit the evidence perfectly. The naive detective despairs, but the master detective realizes this is the most important clue of all! It reveals a hidden symmetry in the problem and tells you exactly what new evidence you need to find to crack the case. In this chapter, we will embark on a journey across the scientific landscape to see this principle in action, from the inner workings of a living cell to the grand sweep of evolutionary history, and even to the halls where decisions about our technological future are made.

Let's begin inside a biological cell, a bustling city of molecular machines. Imagine we are studying a protein, "Proteon," that is being broken down over time. We have a simple model for this: the amount of protein $P(t)$ at time $t$ decays exponentially, like $P(t) = P_0 \exp(-k_d t)$, where $P_0$ is the initial amount and $k_d$ is the degradation rate. We want to find both $P_0$ and $k_d$ by measuring $P(t)$.

Suppose, due to some experimental constraint, we can only start our measurements very late in the process, long after most of the protein has vanished. What happens? We can still observe the rate at which the remaining protein is disappearing, which gives us a decent estimate of the decay constant $k_d$. But how much was there to begin with? It's like arriving at the very end of a party. You can see the rate at which the last few guests are trickling out the door, but you have almost no clue whether it was a wild bash with hundreds of people or a quiet gathering of ten. Any attempt to extrapolate back to the beginning is fraught with huge uncertainty. The initial amount, $P_0$, has become practically non-identifiable due to our choice of experimental design. The lesson is immediate: when you look matters as much as what you look at.

Sometimes, the problem is not in our experimental timing, but is woven into the very fabric of the system and our observation of it. Consider how a drug moves through the body. A simple model might involve a central compartment (the blood) and a peripheral one (the body's tissues). When a drug is injected into the blood, it can be eliminated from the body (with a rate constant $k_{cl}$) or it can move into the tissues (with a rate constant $k_{12}$). If our only measurement is the drug's concentration in the blood, we see the drug disappearing. But is it being eliminated for good, or is it just hiding in the tissues? From the perspective of the blood, both processes contribute to the drug's departure. The equation governing the drug amount in the blood, $A_1$, turns out to depend only on the sum of the rates: $\frac{dA_1}{dt} = -(k_{cl} + k_{12})A_1$.

No matter how perfectly or frequently we measure the blood concentration, we can never separate $k_{cl}$ from $k_{12}$. We can only determine their sum. This is a structural non-identifiability. The two parameters are fundamentally entangled by the structure of the model and our limited view of it. It's like listening to a conversation between two people from outside a closed door; you can measure the total volume of the sound, but you cannot tell how much each person is contributing.

This challenge escalates in more complex systems, like the voltage-gated ion channels that act as the gatekeepers of neural communication. A model of a channel's behavior might involve a complex dance between multiple closed, open, and inactivated states. The macroscopic current we measure is a product of several factors: the number of channels ($N$), the conductance of a single channel ($g$), and the probability that a channel is open ($P_O(t)$). Immediately, we see that we can only identify the product $Ng$, not the individual components. But the problem runs deeper. It's possible for two completely different, intricate choreographies of states—two different Markov models—to produce the exact same probability of being open, $P_O(t)$, over time. From the outside, looking only at the total current, these two distinct mechanisms are perfectly indistinguishable.

Does this mean we give up? No! This is where the detective work gets exciting. We can design a cleverer experiment. For instance, by rapidly changing the voltage, we can catch the channels "off-guard" and measure what are called "tail currents." This new kind of data provides a different view of the system, breaking the ambiguity and allowing us to tell the competing models apart. Science progresses by finding new ways to ask questions that force Nature to give an unambiguous answer. This principle is at the heart of the most advanced biomedical research, such as "systems vaccinology," where scientists sift through immense datasets of genes ($p$ features) and proteins ($q$ features) from a small number of people ($n$ subjects, where $n \ll p,q$) to find the subtle signatures of a successful immune response. Here, naive statistical approaches are guaranteed to find spurious correlations. Principled approaches, like multi-omics factor analysis, are designed with identifiability in mind. They seek the underlying, stable "factors"—coordinated programs of genes and proteins—while using rigorous validation techniques to ensure that what is found is a genuine biological signal and not a statistical ghost born from the high-dimensional noise.

The ghost of non-identifiability also haunts our attempts to reconstruct the deep past. In evolutionary biology, we infer the history of life by comparing the DNA of living organisms. The differences between two sequences are the result of mutations accumulating over time. A key parameter is the branch length of a phylogenetic tree, which represents the evolutionary distance between two species. This distance, however, is a product of the mutation rate ($\mu$) and the divergence time ($t$). From the sequence data alone, we can only estimate their product, $\mu t$. A fast rate over a short time looks identical to a slow rate over a long time.

This is a profound structural non-identifiability. To make any progress, biologists must adopt a convention. They choose to measure branch lengths in a standard unit: the expected number of substitutions per site. This is like agreeing to measure all car journeys in total miles traveled, even if we don't know the exact hours spent driving or the speed in miles per hour. It provides a consistent, albeit relative, measure of evolutionary change.

This very issue lies at the center of one of the great debates in macroevolution: does evolution proceed gradually, or in fits and starts? One model, "gradualism," posits that traits change continuously through time (anagenesis). The expected variance between species would be proportional to the time they have been diverging, described by a covariance matrix $\sigma^2 \mathbf{C}_t$. An alternative model, "punctuated equilibria," suggests that most change happens rapidly during speciation events (cladogenesis). Here, the expected variance would be proportional to the number of branching events shared between species, described by a different matrix $\tau^2 \mathbf{C}_s$.

Which model is correct? Here comes the twist: if the speciation rate is constant over time, then the number of speciation events is, on average, directly proportional to time. This means the two matrices become proportional: $\mathbf{C}_s \approx b \mathbf{C}_t$. If this is the case, then the statistical patterns predicted by the two grand theories of evolution are identical! The gradual and punctuational models become non-identifiable. Nature, by being too regular, hides its preferred mechanism from us. The only way we can hope to distinguish these modes of evolution is to study situations where the tempo of speciation is irregular, breaking the proportionality between the two models.

The same kind of confounding appears in a much more down-to-earth context: counting wildlife. An ecologist wants to know the probability, $\psi$, that a certain species occupies a forest patch. They go out for a survey, but the species might be hard to see. There is a probability, $p$, of detecting it, given it is present. If the ecologist visits a patch and sees the animal, great. But what if they don't? Does it mean the animal isn't there (an occupancy of 0), or that it is there but was missed? The probability of observing the species in a single visit is the product $\psi p$. Are there few animals that are easy to spot (high $p$, low $\psi$), or many animals that are hard to spot (low $p$, high $\psi$)? With just one visit, it's impossible to tell. The parameters are structurally non-identifiable.

But what if the ecologist visits a second time? Suddenly, the case can be cracked. The probability of seeing the animal on both visits is $\psi p^2$. The probability of seeing it on the first visit but not the second is $\psi p(1-p)$. The mathematical forms are different! By comparing the frequencies of these different detection histories, we can solve for $p$ and $\psi$ separately. A simple, brilliant change in experimental design defeats the non-identifiability.

This dialogue between model ambiguity and experimental design is not confined to biology; it is just as crucial in engineering and the physical sciences. Imagine a materials scientist studying creep—the tendency of a metal part to slowly deform under constant stress at high temperature. They observe that the rate of deformation accelerates over time, leading to eventual rupture. Two stories could explain this. One story (Model S) is that the material's intrinsic properties are softening and becoming less resistant. Another story (Model D) is that the material isn't getting weaker, but tiny micro-cracks are forming and growing, which increases the effective stress on the remaining material.

With a single, simple creep test, both stories can be made to fit the data perfectly. They are non-identifiable. To distinguish them, we need to be cleverer. We could use ultrasound to listen for the formation of the micro-cracks, directly testing the premise of Model D. Or, even more subtly, we could give the material a tiny, sudden "kick"—a small, rapid increase in stress—during the experiment. The two models predict a different instantaneous response to this kick. The way the material answers our "kick" reveals its inner nature and tells us which story is true.

Nowhere do these concepts—structural limits, practical uncertainties, and the demand for clever design—come together more powerfully than in the governance of new technologies. Consider a synthetic biologist who has engineered a microbe for release into the environment, perhaps to clean up a pollutant. To get approval, they must demonstrate its safety. They build a mathematical model to predict the microbe's population dynamics and its effect on a native species.

However, their field sensor is imperfect; it reports a single signal that is a weighted sum of the engineered microbe count, $M(t)$, and the native host count, $N(t)$. The weighting factor, $q$, is unknown. Immediately, a structural non-identifiability arises. The equations show that one can get the exact same sensor reading from a world with a large host population and a small weighting factor as from a world with a small host population and a large weighting factor. Without an independent way to calibrate the sensor (i.e., measure $q$) or observe $N(t)$ directly, the model can never tell us the absolute number of native organisms—a critical variable for assessing ecological risk! A responsible regulatory agency, understanding this, should not accept claims about absolute population sizes from this model. They should demand a better experiment or insist that the risk assessment be framed only in terms of quantities the model can identify.

Furthermore, even with a perfect sensor, the team faces practical non-identifiability. If their field trial is too short, the data may contain very little information about the microbe's natural decay rate, $d$. The parameter will be structurally identifiable, but the confidence interval on its estimate will be enormous. This uncertainty has direct policy implications. A regulator might conclude the evidence is insufficient and require a longer study. Or, applying the precautionary principle, they might demand that the risk assessment be performed assuming the worst-case plausible value for the decay rate—that is, assuming the microbe is highly persistent.

This final example lays the stakes bare. Non-identifiability is not an abstract statistical curiosity. It is a central challenge in linking mathematical models to the real world. Acknowledging it is fundamental to responsible research and innovation. It dictates what we can and cannot claim, guides us toward more informative experiments, and provides a rational basis for making decisions in the face of uncertainty. It is, in the end, a formal language for scientific humility.