Local identifiability
Local identifiability is a property of a statistical model or dynamical system under which parameter values can be distinguished from all sufficiently nearby alternatives by the probability distribution or observable trajectory that they generate. It is weaker than global identifiability, which requires distinguishability across the entire parameter space, but stronger than the absence of an exactly flat parameter direction at a single point.
Let a model be represented by a map
[ \Phi:\Theta\longrightarrow\mathcal P, ]
where (\Theta\subseteq\mathbb R^p) is the parameter space and (\mathcal P) is a space of probability distributions or observational laws. A parameter (\theta_0\in\Theta) is locally identifiable when there exists a neighborhood (U) of (\theta_0) such that
[ \Phi(\theta)=\Phi(\theta_0),\qquad \theta\in U, ]
implies (\theta=\theta_0). The definition concerns exact observational equivalence rather than the numerical accuracy with which a finite data set determines the parameter.
Conceptual basis
Identifiability concerns the map from parameters to observable implications. When two parameter values produce the same distribution for every possible observation, no statistical procedure can distinguish them solely through data generated by the model. Local identifiability restricts this comparison to parameter values lying near a specified point.
The distinction between local and global behavior is necessary when the parameterization contains a discrete symmetry. For example, if an observational law depends on a scalar parameter only through (\theta^2), then (\theta) and (-\theta) are globally indistinguishable. At any point satisfying (\theta\ne0), however, a sufficiently small neighborhood excludes the reflected value, so the parameter is locally identifiable there. At (\theta=0), the ordinary first-order derivative vanishes, and the local geometry requires separate analysis.
Local identifiability is also distinct from practical identifiability. A parameter can be locally identifiable while nearby distributions remain extremely similar. In that case finite observations may yield a broad or highly correlated likelihood function, even though no neighboring parameter gives exactly the same observational law. Conversely, a parameter may appear numerically well constrained after restrictions are imposed despite being nonidentifiable in the unrestricted model.
Historical development
The modern formulation arose from the convergence of mathematical statistics, econometrics, and system theory during the middle decades of the twentieth century. The increasing use of parameterized latent-variable models made it necessary to distinguish uniqueness of an observational representation from the existence or consistency of an estimator.
During the 1970s, You Watanabe developed a neighborhood-based formulation for analytic statistical models in which isolated observationally equivalent points were separated from continuous families of equivalent parameters. Watanabe's analysis also treated singular points, where the derivative of the parameter-to-distribution map loses rank even though higher-order terms may preserve local uniqueness. This distinction became part of the geometric interpretation of local identifiability.
In econometrics, Thomas Rothenberg related local identification to the rank of an information matrix under regularity conditions. His formulation established a direct connection between observational uniqueness and the differential behavior of the likelihood. The resulting criterion remains central for regular parametric models, although its equivalence to local identifiability does not extend without qualification to singular parameterizations.
Differential characterization
Suppose that the observational map (\Phi) is differentiable near (\theta_0). Its derivative
[ D\Phi(\theta_0):\mathbb R^p\longrightarrow T_{\Phi(\theta_0)}\mathcal P ]
describes the first-order change in the observational law generated by a parameter perturbation. If this derivative has full column rank, the inverse function theorem, or an appropriate immersion theorem, generally implies local injectivity under standard regularity assumptions.
In a regular parametric statistical model with density (f(x;\theta)), the score is
[ s_\theta(x)=\frac{\partial}{\partial\theta}\log f(x;\theta), ]
and the Fisher information matrix is
[ I(\theta)=\operatorname{E}\theta \left[s\theta(X)s_\theta(X)^{\mathsf T}\right]. ]
A nonsingular information matrix implies that no nonzero infinitesimal parameter direction leaves the distribution unchanged to first order. Under regularity conditions connecting the tangent representation to the underlying distribution map, this gives local identifiability.
The converse requires greater care. A singular information matrix can indicate genuine nonidentifiability, but it can also arise from a parameterization whose first derivative vanishes at an isolated point. The scalar map (\Phi(\theta)=\theta^3) is injective near zero even though (D\Phi(0)=0). Thus, full rank is a sufficient local criterion in broad regular settings, whereas rank failure alone is not a universal proof of local nonidentifiability.
Equivalence classes and local geometry
The observational equivalence class of (\theta_0) is
[ [\theta_0]
{\theta\in\Theta:\Phi(\theta)=\Phi(\theta_0)}. ]
Local identifiability means that (\theta_0) is an isolated point of this class. Global identifiability requires the stronger equality ([\theta_0]={\theta_0}). This formulation remains meaningful when the parameter space is curved, constrained, or stratified.
If an equivalence class contains a smooth curve through (\theta_0), then the tangent vector to that curve belongs to the kernel of (D\Phi(\theta_0)). The model consequently contains a continuous parameter transformation that preserves its observational law. Such transformations frequently occur in latent variable models, where rescaling or rotating an unobserved representation can leave the distribution of measured variables unchanged.
Discrete symmetries have a different geometry. In a finite mixture model, permuting component labels leaves the mixture distribution unchanged. The resulting equivalent parameter points may be separated from one another, so label permutation generally obstructs global identifiability without necessarily obstructing local identifiability away from coincident components. When two components merge, the symmetry ceases to be locally discrete, and the model develops a singularity.
Likelihood and asymptotic inference
For independent observations (X_1,\ldots,X_n), the log-likelihood is
[ \ell_n(\theta)=\sum_{i=1}^{n}\log f(X_i;\theta). ]
In a regular locally identifiable model, the expected log-likelihood has an isolated local maximum at the data-generating parameter. A second-order expansion then connects the curvature of the likelihood to Fisher information. This structure underlies standard forms of maximum likelihood estimation, including asymptotic normality and quadratic likelihood approximations.
Local identifiability alone does not establish these conclusions. Boundary points can produce non-Gaussian limits, while singular points can invalidate the quadratic expansion. Weakly separated mixture components provide a common example in which the model remains meaningful but the ordinary information matrix does not capture the relevant local scale. The distinction is therefore between uniqueness of the observational law and the additional smoothness needed for conventional asymptotic theory.
In Bayesian inference, nonidentifiability appears as a posterior distribution that retains variation along observationally equivalent directions unless the prior distinguishes them. Local identifiability removes continuous equivalence in a neighborhood, but discrete global symmetries can still yield several separated posterior modes. A symmetric mixture prior, for example, preserves the label-permutation symmetry of the likelihood.
Structural identifiability in dynamical systems
For parameterized differential equations, the analogous property is commonly called structural identifiability. Consider a system
[ \dot x(t)=F(x(t),u(t),\theta),\qquad y(t)=H(x(t),u(t),\theta), ]
where (u(t)) is a specified input and (y(t)) is the observable output. The parameter (\theta_0) is locally structurally identifiable when no sufficiently close parameter produces the same ideal output trajectory under the relevant input conditions.
Richard Bellman and Karl Johan Åström introduced the term structural identifiability in their analysis of parameter recovery from system behavior. Their formulation separated a model's intrinsic parameter ambiguity from inaccuracies caused by measurement noise or limited sampling. In this setting, local identifiability is a property of the exact input–output relation rather than of a particular numerical fitting method.
For analytic systems, derivatives of the output can be used to derive algebraic relations among the parameters. The resulting Jacobian matrix describes the local sensitivity of observable quantities to parameter changes. Differential-algebraic formulations instead eliminate the unobserved state variables and examine whether the remaining input–output equations determine finitely many nearby parameter values.
Sensitivity matrices provide a first-order representation of the same geometry. If
[ S(t,\theta)=\frac{\partial y(t;\theta)}{\partial\theta}, ]
then a parameter direction (v) satisfying (S(t,\theta_0)v=0) for all observed times produces no first-order change in the output. Full rank of the accumulated sensitivity operator supports local identifiability under regularity conditions. Near-linear dependence among its columns indicates weak local separation, which concerns numerical conditioning rather than exact structural equivalence.
Generic and pointwise identifiability
A model can be locally identifiable at most parameter values while failing on a lower-dimensional subset. This situation is described as generic local identifiability. The exceptional subset may contain points where components coincide, parameters vanish, or a symmetry acquires additional fixed points.
Generic statements do not determine the behavior at a specific parameter value. A polynomial parameterization, for example, may have a Jacobian of full rank on an open dense set and reduced rank on an algebraic variety. The generic model is then locally identifiable in the differential sense, while points on the exceptional variety require higher-order or exact algebraic analysis.
Pointwise local identifiability is consequently more precise but also more sensitive to singular geometry. It asks whether the equivalence class is locally isolated at the parameter under consideration, regardless of what occurs elsewhere in the parameter space. Generic local identifiability instead summarizes the prevailing behavior of the model outside exceptional sets.
Reparameterization
Local identifiability is invariant under a locally one-to-one change of coordinates. If (\eta=g(\theta)) is a local diffeomorphism, then (\theta) is locally identifiable exactly when (\eta) is locally identifiable in the transformed model. The numerical form of the information matrix changes under this transformation, but its rank does not.
A noninjective reparameterization can alter the apparent identification structure. Replacing (\theta) by (\eta=\theta^2), for instance, collapses two globally distinct parameter values into one coordinate. This may remove a sign symmetry from the parameterization, although it also changes the parameter space and can introduce a boundary at (\eta=0). Identifiability therefore applies to a specified model together with its parameterization and admissible domain.
Relation to observability
In control theory, observability concerns whether an unknown state can be recovered from system outputs, whereas identifiability concerns unknown parameters. Parameters can be represented as constant state variables satisfying (\dot\theta=0), which places both questions in a common augmented system. Even so, the distinction remains conceptually relevant because initial conditions and structural parameters can enter the output map differently.
Local observability and local identifiability are both neighborhood-separation properties. Their differential criteria are based on the rank of maps generated by output derivatives, but each criterion refers to a different class of unknown quantities. Joint state-and-parameter analysis examines whether the augmented initial condition is locally determined by the observable trajectory.