Computational Identifiability: Concepts & Methods

• Computational identifiability is the study of uniquely recovering latent parameters from observable data through concrete algorithms, addressing issues like injectivity and stability.
• It leverages algebraic, numerical, and graphical methods to analyze parameter-to-observable maps across diverse models such as ODEs, PDEs, and causal frameworks.
• Recent advances in symbolic algorithms, rank criteria, and structured optimization demonstrate how theory and computational feasibility converge in solving inverse problems.

Searching arXiv for papers on computational identifiability and closely related formulations. arXiv search query: "computational identifiability" Computational identifiability studies whether latent model quantities can be uniquely, finitely, or stably recovered from observable data, and how that determination can be carried out by concrete algorithms rather than by abstract existence arguments alone. In the cited literature, the observable object may be a covariance matrix, a conditional distribution, an input-output trajectory, or a noisy distributed observation field, while the quantities to be identified may be sparse drift matrices, structural coefficients, representation functions, causal effects, Hamiltonian parameters, or spatially varying fields (Dettling et al., 2022, Dörfler et al., 2024, Roeder et al., 2020, Bakeer, 27 May 2026). The subject therefore spans algebraic injectivity, rank criteria, symbolic elimination, semialgebraic decision procedures, and local information geometry, with a recurring distinction between what is identifiable in principle and what is computationally feasible to decide or recover.

1. Meanings and scope of computational identifiability

The literature does not use a single universal notion of identifiability. Instead, it distinguishes several related properties according to the model class and the computational question being asked. A model is globally identifiable when the parameter-to-observable map is injective everywhere on the admissible parameter domain; it is generically identifiable when injectivity holds outside a proper algebraic subset; it is numerically identifiable when, for a specific feasible observed object such as a covariance matrix, there is exactly one parameter explanation; and it is practically identifiable when small perturbations in observations lead to small perturbations in the fitted parameter (Dörfler et al., 2024, Wang et al., 2 Jan 2025).

A separate but important shift appears in deep representation learning. There, classical parameter-space identifiability is typically absent because different neural-network parameter vectors can represent the same function, yet the learned representation functions can still be identifiable in function space up to an invertible linear transformation. In the canonical discriminative family

$$p(Y \mid X, S) \;=\; \frac{\exp\!\big(f(X)^\top g(Y)\big)} {\sum_{Y' \in S} \exp\!\big(f(X)^\top g(Y')\big)},$$

the relevant object of identification is therefore the pair $(f,g)$, not the underlying network weights, and the resulting notion is linear identifiability rather than parameter uniqueness (Roeder et al., 2020).

In systems biology and inverse problems, the distinction between structural identifiability and practical identifiability remains central. Structural identifiability asks whether idealized noiseless data determine the parameter map injectively, whereas practical identifiability asks whether finite, noisy, and design-dependent data constrain parameters with useful precision. This separation is explicit in the review literature and is extended to PDE and distributed-parameter settings, where local visibility and pattern distinguishability replace purely scalar notions of recoverability (Preston et al., 26 Aug 2025, Renardy et al., 2021).

Notion	Core criterion	Representative setting
Global identifiability	injective everywhere	continuous Lyapunov models, linear SCMs
Generic identifiability	injective off a proper algebraic subset	linear SCMs, latent-variable Gaussian graph models
Numerical identifiability	unique fiber for a given feasible $\Sigma$	linear SCMs
Practical identifiability	perturbation-stable recovery; FIM invertibility	biological model fitting
Linear identifiability	equality of conditional distributions implies equivalence up to invertible linear maps	deep representation learning

This diversity of notions suggests that “computational identifiability” is best understood as a family of inverse problems tied together by a common theme: the observable data define a fiber in parameter space, and the computational task is to determine the size, structure, or local geometry of that fiber.

2. Parameter-to-observable maps, fibers, and algebraic reduction

A common formal pattern is to represent the model by a parameter-to-observable map and study its fibers. In continuous Lyapunov models, the latent parameter is a sparse drift matrix $M$, the observable is the stationary covariance $\Sigma$, and the relation is the continuous Lyapunov equation

$M\Sigma + \Sigma M^\top + C = 0.$

Fixing a graph $G$ that specifies the zero pattern of $M$, the parameter-to-covariance map is the rational map

$$\phi_{G,C}:\mathrm{Stab}(E)\to PD_p,\qquad M\mapsto \Sigma(M,C),$$

and identifiability becomes an injectivity question for $\phi_{G,C}$ or, equivalently, uniqueness of solutions of the Lyapunov equation inside the support-constrained class $(f,g)$0 (Dettling et al., 2022).

In linear structural causal models, the analogous observable is again a covariance matrix,

$(f,g)$1

with unknown directed coefficients $(f,g)$2 and noise covariance $(f,g)$3 constrained by an acyclic mixed graph $(f,g)$4. The fiber

$(f,g)$5

collects all parameter pairs producing the same covariance. Global, generic, and numerical identifiability then amount to different cardinality statements about this fiber, and the central computational move is to encode the non-identifiable locus as a semialgebraic set whose dimension can be compared with the ambient parameter space (Dörfler et al., 2024).

In ODE models, the corresponding reduction goes through differential algebra. State variables are eliminated to produce input-output equations, and the field of identifiable parameter functions is expressed as a field intersection. For a rational ODE model $(f,g)$6, the field of single-experiment identifiable functions is

$(f,g)$7

and the paper “Computing all identifiable functions of parameters for ODE models” shows that this field can be computed as

$(f,g)$8

where $(f,g)$9 is built from Wronskians of monomials in the input-output equations, without requiring the traditional solvability condition (Ovchinnikov et al., 2020).

Several model classes then reduce identifiability to explicit linear algebra. In continuous Lyapunov models, solving for $\Sigma$0 yields a matrix equation

$\Sigma$1

and global or generic identifiability becomes a full-column-rank condition on $\Sigma$2. The same paper gives an equivalent kernel formulation using a smaller matrix $\Sigma$3, where identifiability is equivalent to full column rank of $\Sigma$4 (Dettling et al., 2022). In linear compartmental models, structural identifiability is similarly tested through the Jacobian of a coefficient map built from reduced input-output equations, and generic local identifiability holds exactly when that Jacobian has rank equal to the number of free parameters (Gross et al., 2018).

A different algebraic reduction appears in discrete-parameter phylogenetic models. There the relevant objects are model varieties, and generic identifiability of the discrete topology parameter is certified by finding a coordinate subset that is independent in the algebraic matroid of one model variety but dependent in another. Jacobian rank computations replace direct Gröbner-basis computations of full vanishing ideals, which is the main computational innovation of that work (Hollering et al., 2019).

3. Graphical and structural criteria

In many settings, identifiability is governed by graph structure rather than by generic dimension counting alone. Continuous Lyapunov models provide one of the cleanest examples. If sparsity in the drift matrix $\Sigma$5 is encoded by a directed graph $\Sigma$6 with self-loops, then a graph is called simple when it contains no directed two-cycle. The main theorem states: $\Sigma$7 Moreover, if $\Sigma$8 is diagonal, then

$\Sigma$9

Thus feedback cycles of length at least $M$0 remain compatible with global identifiability, while reciprocal pairs are the exact obstruction in the diagonal-noise case (Dettling et al., 2022).

For non-simple graphs in the same model family, generic identifiability is subtler. The paper introduces a necessary trek-based inequality,

$M$1

but also gives explicit counterexamples showing that this condition is not sufficient, because hidden symbolic linear dependencies can still force non-identifiability (Dettling et al., 2022). This suggests that graph separation and parameter counting alone do not exhaust the computational content of identifiability.

Graph criteria also organize identifiability in latent-variable Gaussian models. For directed Gaussian graphical models with one latent source, the problem is not global injectivity but generic finite-to-one identifiability. The sufficient graph criterion is that every connected component of the undirected complement $M$2 contains an odd cycle. The necessary condition sharpens this through two derived graphs, $M$3 and $M$4, and again counts complement components without odd cycles. The same paper also proves that generic finite identifiability is invariant under Markov equivalence and gives a subgraph extension principle based on deleting certain sink or source nodes (Leung et al., 2015).

Partial causal knowledge leads to a different kind of graph-based hierarchy. In causal abstractions formalized as classes $M$5 of compatible causal diagrams, the paper “Identifiability in Causal Abstractions: A Hierarchy of Criteria” distinguishes Identifiability through Graphs (IG), Identifiability through Graphs knowing $M$6 (IGP), Identifiability by Common Do-Calculus (ICD), and Identifiability by Common Graphical Criterion (ICGC), together with specific cases such as common backdoor or frontdoor criteria. The resulting hierarchy is

$M$7

and the paper proves that identifiability over the full class $M$8 reduces to checking only the maximal graphs $M$9 under inclusion (Yvernes et al., 8 Jul 2025).

A related structural reduction appears in linear compartmental models. The input-output equation for a given measured compartment can be computed on the output-reachable subgraph rather than on the full graph, and parameters outside the observable component do not affect the reduced coefficient map. This reduction is important because identifiability can otherwise be misread from higher-order input-output consequences that are not reduced enough for coefficient-map analysis (Gross et al., 2018).

4. Symbolic algorithms and computational complexity

A major theme in computational identifiability is the separation between abstract decidability and feasible symbolic computation. In linear SCMs, the paper “On the Complexity of Identification in Linear Structural Causal Models” gives a complete decision procedure for generic identifiability in polynomial space by encoding the non-identifiable locus

$\Sigma$0

as a semialgebraic set and comparing its dimension with that of the parameter space. By Renegar’s algorithm for the theory of the reals with a constant number of quantifier alternations, generic identifiability is in $\Sigma$1, and hence in at most exponential time, improving on a prior complete Gröbner-basis approach with double-exponential worst-case time (Dörfler et al., 2024).

The same paper also establishes the first hardness result in this area for a notion of identifiability. The promise problem of deciding, for a feasible covariance matrix $\Sigma$2, whether the fiber has exactly one parameter explanation is proved $\Sigma$3-hard, and therefore $\Sigma$4-hard. Thus exact uniqueness checking for a concrete observed covariance matrix is computationally hard even though generic identifiability is decidable in PSPACE (Dörfler et al., 2024). This sharpens the distinction between property-level identifiability and algorithmic tractability.

For rational identification formulas in linear SCMs with latent confounding, a different symbolic strategy is now available. “Efficient Symbolic Computations for Identifying Causal Effects” replaces full nonhomogeneous Gröbner-basis elimination by a weighted-homogeneous formulation derived from maximal trek lengths. If a causal effect admits an identification formula of maximal degree $\Sigma$5, the proposed algorithm searches degree by degree in a weighted homogenized ideal and provably finds the lowest-degree identifying formulas within that bound. The runtime is quasi-polynomial in graph size for fixed $\Sigma$6, and the method substantially outperforms the standard García-Puente Gröbner-basis procedure on random $\Sigma$7-node instances (Hollering et al., 22 Apr 2026).

ODE identifiability admits another exact symbolic solution. “Computing all identifiable functions of parameters for ODE models” gives the first algorithm for computing the entire field of single-experiment identifiable rational functions without any solvability assumption. It also proves that the field generated by the coefficients of input-output equations is exactly the field of functions identifiable from sufficiently many generic experiments, and gives a sufficient experiment bound

$\Sigma$8

where $\Sigma$9 and $M\Sigma + \Sigma M^\top + C = 0.$0 come from a Wronskian rank calculation on the input-output equations (Ovchinnikov et al., 2020).

Algebraic matroids provide a further symbolic shortcut for discrete parameters. In phylogenetic mixtures and networks, the paper “Identifiability in Phylogenetics using Algebraic Matroids” replaces full ideal computation by Jacobian-based matroid separation and proves generic identifiability of 2-tree CFN and K3P mixtures and of $M\Sigma + \Sigma M^\top + C = 0.$1-cycle network parameters under K2P and K3P models. The method is explicitly sufficient rather than complete, but it reduces reliance on Gröbner bases in large families of cases (Hollering et al., 2019).

Other model classes translate identifiability into structured optimization rather than elimination. In mixtures of linear chirps sampled noiselessly,

$M\Sigma + \Sigma M^\top + C = 0.$2

the paper “Pursuing the limit of chirp parameter identifiability: A computational approach” proves that $M\Sigma + \Sigma M^\top + C = 0.$3 is necessary for $M\Sigma + \Sigma M^\top + C = 0.$4, and shows that unique identifiability is equivalent to uniqueness of a structured rank-constrained matrix feasibility problem involving a lifted matrix $M\Sigma + \Sigma M^\top + C = 0.$5, a multilevel Toeplitz matrix $M\Sigma + \Sigma M^\top + C = 0.$6, a PSD block constraint, and the interpolation conditions $M\Sigma + \Sigma M^\top + C = 0.$7 (Yang et al., 2 Jul 2025). This is an example of computational identifiability being characterized by uniqueness of a nonconvex but highly structured optimization problem.

5. Local, practical, and spatial identifiability

Not all identifiability questions are global or symbolic. In distributed-parameter inverse problems, the relevant issue is often which perturbation patterns are locally distinguishable under a given sensing and excitation program. “Local Information Operators for Spatial Identifiability in Distributed-Parameter Inverse Problems in Computational Mechanics” develops this viewpoint around the operator

$M\Sigma + \Sigma M^\top + C = 0.$8

or, in Hilbert-space notation, $M\Sigma + \Sigma M^\top + C = 0.$9. For locally linearized Gaussian models with parameter-independent covariance, this operator is simultaneously Fisher information, Gauss–Newton data-misfit curvature, and a noise-weighted sensitivity Gramian. Its diagonal $G$0 gives a pointwise information density, while its kernel and spectra identify strongly visible, weakly visible, and locally invisible spatial patterns (Bakeer, 27 May 2026).

This local-operator perspective separates pointwise visibility from full spatial identifiability. The nullspace satisfies

$G$1

so perturbations $G$2 with $G$3 are locally unobservable. Prior information, nuisance parameters, and model discrepancy enter the same geometry through

$G$4

through covariance inflation $G$5, and through the Schur-complement information loss

$G$6

which quantifies confounding by nuisance parameters (Bakeer, 27 May 2026). A plausible implication is that in large inverse problems, identifiability is often better understood as a geometry of informed subspaces than as a binary parameter-level property.

Practical identifiability in biological model fitting is formalized differently but leads to a similar local conclusion. “A Systematic Computational Framework for Practical Identifiability Analysis in Mathematical Models Arising from Biology” defines practical identifiability as perturbation-stable recovery of the least-squares estimator and proves: $G$7 With sensitivity matrix $G$8,

$G$9

and the paper also proves that coordinate identifiability is equivalent to invertibility of the same FIM. When the FIM is singular, coordinate non-identifiability of parameter $M$0 is characterized by whether the sensitivity column $M$1 lies in the span of the remaining columns, leading to the computational metric

$M$2

as a fast linear approximation to profile likelihood (Wang et al., 2 Jan 2025).

The systems-biology review literature places these local tools in a broader hierarchy. For linear Gaussian models, the Fisher information matrix gives an exact characterization of uncertainty, and in nonlinear models the local approximation

$M$3

links sensitivity, sloppiness, and practical recoverability. Profile likelihood, global sensitivity analysis, synthetic-data recovery experiments, and MCMC then probe departures from the local Gaussian picture, especially when multiple minima or parameter symmetries are present (Preston et al., 26 Aug 2025).

Age-structured PDE models show how these ideas extend beyond ODEs. The paper on structural identifiability of age-structured epidemic PDEs constructs input-output differential polynomials in both $M$4 and $M$5, augments them with boundary equations at $M$6, and compares the resulting identifiable combinations with those of corresponding ODE systems. In the constant-parameter SEI models studied there, the PDE and ODE often have the same identifiable combinations, but age-resolved immigration and age-dependent death rates create strictly richer identifiability structure (Renardy et al., 2021).

6. Domains, limitations, and open questions

Computational identifiability now appears across a wide range of domains. In continuous Lyapunov models it governs recovery of sparse drift matrices from equilibrium covariances (Dettling et al., 2022). In linear and mixed graphical causal models it governs recovery of structural coefficients and causal effects from observational covariances (Dörfler et al., 2024, Hollering et al., 22 Apr 2026). In dynamical biology it governs parameter combinations recoverable from trajectories, input-output equations, and multiple experiments (Ovchinnikov et al., 2020, Preston et al., 26 Aug 2025). In distributed mechanics it governs which spatial field perturbations are visible under specified sensors and load programs (Bakeer, 27 May 2026). In deterministic signal models it becomes uniqueness of chirp decomposition from regularly spaced noiseless samples (Yang et al., 2 Jul 2025). In quantum systems it becomes structural global identifiability of Hamiltonian coefficients and, beyond that, the existence of “economic” identification algorithms whose complexity depends on the number of unknown parameters rather than directly on Hilbert-space dimension (Wang et al., 2018).

Several recurring limitations are equally clear. First, many results are local or generic rather than global. Practical identifiability via FIM invertibility is local by construction (Wang et al., 2 Jan 2025), and the local information operator for distributed-parameter inverse problems diagnoses first-order distinguishability around a nominal field rather than global uniqueness (Bakeer, 27 May 2026). Second, graph criteria are often sufficient or necessary but not complete: trek inequalities in continuous Lyapunov models are necessary but not sufficient (Dettling et al., 2022), and algebraic matroids may fail to distinguish models with different ideals but the same coordinate-dependence structure (Hollering et al., 2019). Third, exact symbolic decidability can remain computationally formidable even when the mathematical property is well defined, as illustrated by $M$7-hardness for numerical identifiability in linear SCMs (Dörfler et al., 2024).

Open problems therefore concentrate on the gap between exact theory and effective computation. The continuous Lyapunov literature explicitly points to improved understanding of generic identifiability for non-simple graphs and to broader symbolic certification methods for larger graphs (Dettling et al., 2022). The causal-abstraction literature leaves open whether IG can hold without ICD, that is, whether a uniform estimand can exist over a graph class without a common do-calculus proof (Yvernes et al., 8 Jul 2025). The degree-bounded symbolic approach for causal effects leaves open how the maximal degree of identifying formulas grows with graph size and structure (Hollering et al., 22 Apr 2026). Systems-biology reviews continue to emphasize that good fit alone is not enough: non-identifiable or weakly identifiable directions can remain harmless in the calibration regime but become consequential for extrapolative predictions (Preston et al., 26 Aug 2025).

Taken together, these developments show that computational identifiability is no longer a narrow preprocessing step. It is a general program for turning inverse uniqueness questions into concrete rank tests, semialgebraic decision problems, structured optimization tasks, or local information operators, and for clarifying exactly which aspects of a latent model are recoverable from a given observational design.