Generic Rank Tests for Structural Identifiability

Why Rank Enters Structural Identifiability

Structural identifiability asks whether ideal, noise-free observations can distinguish parameter values, given a specified model and its assumptions. A parameter may affect an output only weakly, making it difficult to estimate from real data; that is a practical concern. The structural question is sharper: can two nearby parameter values produce exactly the same ideal output? A derivative map gives a local test for this question by measuring how each parameter changes the output.

Here the model structure, inputs, and initial conditions are treated as specified. The unknown parameter vector is \(\theta=(\theta_1,\ldots,\theta_p)\) in an open parameter domain \(\Omega\). The ideal output is assumed to be differentiable with respect to \(\theta\). In analytic models, the output map is often analytic as well, which supports the generic-rank conclusions below. These assumptions matter: changing which initial conditions or inputs are known changes the map and can change its rank.

Definition: For observations of one or more outputs at times \(t_1,\ldots,t_m\), the parameter-to-output map is \(G:\Omega\to\mathbb{R}^m\), with components \(G_i(\theta)=y(t_i;\theta)\) for a single output. Its derivative matrix, or Jacobian, is \(J(\theta)\), where \(J_{ij}(\theta)=\partial y(t_i;\theta)/\partial\theta_j\). For a continuous ideal output, the derivative map sends a parameter perturbation \(h\) to the output perturbation \(\sum_{j=1}^p [\partial y(t;\theta)/\partial\theta_j]h_j\), considered over the observed time interval.

Each column of a finite-observation Jacobian describes the local output change produced by changing one parameter. If the columns are linearly independent, no nonzero infinitesimal parameter change leaves all the recorded ideal outputs unchanged to first order. For a continuous trajectory, the corresponding condition is that no nonzero \(h\) makes the sensitivity combination equal to zero at every observed time.

Generic Rank and the Local Test

The rank at one parameter value is not always enough to characterize the model. A Jacobian can lose rank at special parameter values even though it has full column rank almost everywhere. The generic rank is the largest rank attained on the parameter domain. In an analytic model, the set where a nonzero maximal minor vanishes is exceptional; away from such rank-loss points, the Jacobian has its generic rank.

Theorem: Suppose the parameter-to-output map is continuously differentiable on an open parameter domain. If its Jacobian has column rank \(p\) at \(\theta_0\), then the parameter is locally identifiable at \(\theta_0\) from those outputs: sufficiently nearby parameter values cannot produce the same outputs. Thus, full column rank gives local identifiability at every point where it holds. If the map is analytic on a connected domain and has generic rank \(p\), the full-rank points are generic because a nonzero \(p\)-by-\(p\) minor cannot vanish on an open subset. If the map is analytic and its generic rank is \(r<p\), then at generic regular points the constant-rank theorem gives local sets of indistinguishable parameter values of dimension \(p-r\). Rank deficiency at a single exceptional point, by itself, does not establish nonidentifiability there.

“Locally identifiable” means that the parameter value is isolated among parameter values giving the same ideal output, in some neighborhood. It does not mean that it is the only such value in the entire parameter domain. In particular, a model can be locally identifiable at a generic point and still have a separate, distant parameter value with the same output.

Why Full Rank Implies Local Identifiability

For a finite observation map, a Jacobian with column rank \(p\) has a nonzero \(p\)-by-\(p\) minor. Select the \(p\) output components corresponding to that minor and call their submap \(H\). Its square Jacobian has nonzero determinant at the parameter value. The inverse function theorem then says that \(H\) is one-to-one in a sufficiently small neighborhood. If two nearby parameter values gave the same full output vector \(G\), they would also give the same selected components \(H\), which is impossible unless the parameter values were equal. This proves local identifiability.

Conversely, if the generic rank is \(r<p\), then at a regular point where rank is locally constant, the constant-rank theorem describes the nearby level set of the output map as having dimension \(p-r\). Since \(p-r>0\), this level set contains distinct, arbitrarily nearby parameter values with identical outputs. This is the local source of generic nonidentifiability. It is a statement about regular points, not a rule that every rank-deficient singular point must be nonidentifiable.

Worked Example: Two Observations of Exponential Decay

Consider the hypothetical model \(\dot{x}=-kx\), with unknown initial concentration \(x(0)=A\). Both \(A\) and \(k\) are unknown, and the ideal output is \(y(t)=A e^{-kt}\). Assume \(A>0\), \(k>0\), and observations are made at \(t=1\) and \(t=3\) time units. The parameter-to-output map is \(G(A,k)=(A e^{-k},A e^{-3k})\).

Differentiating each output with respect to \(A\) and \(k\) gives

J(A,k)= \begin{pmatrix} e^{-k} & -A e^{-k}\\ e^{-3k} & -3A e^{-3k} \end{pmatrix}.

Its determinant is

\det J = -3A e^{-4k}+A e^{-4k} = -2A e^{-4k}.

For every \(A>0\) and finite \(k>0\), this determinant is nonzero, so the map has generic rank 2 and both parameters are locally structurally identifiable from these two ideal observations. For a numerical check, take the hypothetical values \(A=10\) concentration units and \(k=0.2\) per time unit. Then

J(10,0.2)\approx \begin{pmatrix} 0.8187 & -8.187\\ 0.5488 & -16.46 \end{pmatrix}, \qquad \det J=-20e^{-0.8}\approx -8.987.

The first column has units of output per concentration unit of \(A\); the second has units of output per unit of \(k\). The nonzero determinant establishes local distinguishability for this observation schedule. It does not alone establish that no other, distant parameter pair could give the same observations.

Worked Example: One Observation Cannot Separate Two Decay Parameters

Keep the same model and unknowns, but observe only once at \(t=2\) time units. The map is \(G(A,k)=A e^{-2k}\), so its Jacobian is the one-row matrix

J(A,k)= \begin{pmatrix} e^{-2k} & -2A e^{-2k} \end{pmatrix}.

This row is nonzero when \(A>0\), so the rank is 1, less than the two unknown parameters. For a concrete hypothetical observation, take \(A=10\) concentration units and \(k=0.2\) per time unit. The output is \(10e^{-0.4}\approx 6.703\) concentration units. For any nearby \(k'\), choosing \(A'=10e^{2(k'-0.2)}\) preserves the exact original output. Thus a continuum of nearby parameter pairs matches the single ideal observation. Neither parameter can be separately identified from that observation, even with no measurement noise.

Generic Rank Can Hide Exceptional Points and Global Ambiguity

Worked Example: Full Generic Rank With an Exceptional Point

Consider the hypothetical two-output map \(G(\theta_1,\theta_2)=(\theta_1+\theta_2,\theta_1\theta_2)\), with positive parameters. Its Jacobian and determinant are

J(\theta_1,\theta_2)= \begin{pmatrix} 1&1\\ \theta_2&\theta_1 \end{pmatrix}, \qquad \det J=\theta_1-\theta_2.

The generic rank is 2 because the determinant is nonzero whenever \(\theta_1\ne\theta_2\). At the hypothetical values \((\theta_1,\theta_2)=(2,5)\), the outputs are \((7,10)\) and the determinant is \(2-5=-3\), so the parameters are locally identifiable there. However, swapping the parameters gives \((5,2)\), which has the same outputs. The model is not globally identifiable over the unrestricted positive parameter domain.

At \((3,3)\), the Jacobian has rank 1, but the point is still locally identifiable: the only positive parameter pair giving outputs \((6,9)\) is \((3,3)\). The nearby swapped pairs give identical outputs to each other, but those outputs are \((6,9-\epsilon^2)\), not \((6,9)\).

Constructing and Checking the Derivative Map

For a finite set of ideal observations, the construction is direct: state which model outputs are observed, specify their times and known inputs, form the vector of outputs, then differentiate each component with respect to each unknown parameter. If the ideal information is an entire output trajectory, the sensitivity functions are \(S_j(t)=\partial y(t;\theta)/\partial\theta_j\). A parameter direction \(h\) is invisible to first order when

\sum_{j=1}^p S_j(t)h_j=0 \quad\text{for every observed }t.

Independence of the sensitivity functions is the continuous-output counterpart of full column rank. Evaluating sensitivities at selected times produces a finite matrix, but a finite set of times may miss information present in the full trajectory. Therefore, a rank calculation for selected observations establishes a result for that observation design; it does not automatically establish the rank of the entire ideal trajectory map.

In practice, check the derivative calculation and the assumptions behind it before interpreting a rank result. Confirm that the rows correspond to genuinely observed outputs and that known inputs, initial conditions, and parameter constraints match the model specification. Symbolic minors can establish whether rank is generically full. Numerical evaluation at a few parameter values can help diagnose the calculation, but it cannot prove generic rank: a chosen value may lie on an exceptional set. Near-zero numerical singular values also require care because parameter units and scaling affect numerical conditioning, even though exact rank is unchanged by nonsingular rescaling.

What the Test Does—and Does Not—Establish

Approach	What it examines	Key qualification
Generic derivative-rank test	Local distinguishability through sensitivities of ideal outputs	Full generic rank gives generic local identifiability, not global uniqueness
Input-output differential equations	Relations among outputs, their derivatives, inputs, and parameters	Can expose parameter combinations, but requires correct treatment of the assumed ideal relations
Symmetry or invariance analysis	Parameter transformations that leave outputs unchanged	Can demonstrate exact ambiguity, including ambiguities not evident from a local calculation alone

The rank test is most informative when the model and ideal observation specification are fixed. A deficient generic rank identifies locally invisible parameter directions under those assumptions, but changing the inputs, observed outputs, or known initial conditions changes the derivative map. A full-rank result says nothing about how accurately parameters can be estimated with finite, noisy data. Near-dependence among sensitivity columns can make estimation unstable despite exact full rank. Conversely, a poor numerical fit or a small estimated standard error is not a substitute for a structural rank argument.

Common Pitfalls

Treating a single rank-deficient point as proof of generic failure. A determinant can vanish only at an exceptional parameter value. Find the generic rank, for example by examining whether a maximal minor is identically zero.
Calling full rank global identifiability. Full rank gives local uniqueness near regular parameter values. Distinct, distant parameter values may still give the same outputs, as the parameter-swapping example shows.
Confusing structural and practical conclusions. Exact rank concerns ideal outputs. Small sensitivity magnitudes or nearly dependent sensitivity columns concern numerical conditioning and practical estimation.
Using a contradictory model specification. If an initial condition is an unknown parameter, state it as parameter-dependent; do not also treat it as known.

Key takeaway: Build the derivative map from the specified ideal outputs and unknown parameters. Generic column rank equal to the number of parameters supports generic local structural identifiability; lower generic rank signals locally indistinguishable parameter directions at generic regular points. Neither conclusion, by itself, settles global uniqueness or practical estimability.

Check Your Understanding

For each question, first make your own rank or identifiability argument, then compare with the solutions.

A smooth map has three unknown parameters and generic Jacobian rank 3. What local conclusion follows, and what does not follow?
A two-parameter model has generic rank 1 at regular points. What is the local dimension of the indistinguishable parameter sets under the constant-rank condition?
Why can a finite sensitivity matrix have lower rank than the derivative map for an entire ideal trajectory?
For \(G(\theta_1,\theta_2)=(\theta_1+\theta_2,\theta_1\theta_2)\), what is the generic rank, and why does that not imply global identifiability?

Solutions. (1) Full column rank gives local identifiability at generic regular parameter values. It does not rule out distant parameter values with the same outputs. (2) The local indistinguishable sets have dimension \(2-1=1\) at regular points. (3) Selected times may fail to capture sensitivity directions that are distinguishable elsewhere in the trajectory. (4) The determinant is \(\theta_1-\theta_2\), so the generic rank is 2. Swapping the two parameters leaves both the sum and product unchanged, giving a distinct parameter pair with the same outputs.