Imagine trying to deduce the exact shape of a bowl solely by watching a marble roll inside it. If you drop the marble precisely into the center, it stays perfectly still; you learn nothing about the bowl's curvature. If you release it from a highly specific angle so that it rolls in a flat, repetitive circle, you only learn about that single cross-section. To map the entire bowl, you must drop the marble from a "generic" position that forces its trajectory to sweep across the entire surface.
This is the core intuition behind identifiability in dynamical systems—the ability to uniquely determine a system's underlying parameters (such as physical forces, causal structures, or drift matrices) from observed data. A single trajectory is a window into the system's physics, but if that trajectory is poorly initiated, the true governing equations remain hidden. As Yanbing Wang et al. demonstrate in Identifiability of car-following dynamics, theoretical identifiability under ideal conditions does not guarantee practical success; the specific initial conditions and input trajectories dictate whether the observed data contains enough information to uniquely isolate driver behavior.
The Trap of Invariant Subspaces
Why do certain starting points fail to reveal a system's dynamics? The answer lies in the geometry of the state space, specifically invariant subspaces. An invariant subspace is a region of the state space that traps trajectories: if a system starts inside it, its evolution is entirely confined to that region.
Xi Geng et al. show in Identifiability Analysis of Linear ODE Systems with Hidden Confounders that recovering causal parameters requires the system's trajectory to sufficiently span the state space. This spanning behavior is guaranteed when initial conditions are "generic"—meaning they do not reside within restricted mathematical subspaces.
Conversely, if the initial state is confined to a proper invariant subspace of the system's drift matrix, the resulting trajectories fail to explore the full state space. Yuanyuan Wang et al. prove this constraint rigorously for linear stochastic differential equations (SDEs) in Generator Identification for Linear SDEs with Additive and Multiplicative Noise. When a trajectory is trapped within these subspaces, critical components of the governing generator remain completely unobserved.
Uninformative Initial State Generic Initial State
(In Invariant Subspace) (Spans the State Space)
┌───────────┐ ┌───────────┐
│ ┌───────┐ │ │ ───────┐ │
│ │ Trapped │ │ │ / │ │
│ └─►─────┘ │ │ │ ◄───┐ │ │
│ │ │ └───►─┘ │ │
└───────────┘ └───────────┘
Explores restricted subspace. Explores full state space.
Parameters are ambiguous. Parameters are identifiable.
Even when a system is theoretically identifiable from a single trajectory, Cecilia Casolo et al. reveal a major practical bottleneck in Identifiability Challenges in Sparse Linear Ordinary Differential Equations. In the sparse regime—where most variables do not interact—the authors quantify practical unidentifiability by measuring the geometric distance between the initial state and the system's invariant subspaces. If the starting condition is too close to these subspaces, the trajectory yields ambiguous data that perfectly fits multiple competing model candidates, rendering the true system unrecoverable.
Generating Geometric Diversity
To guarantee identifiability, the system's trajectories must exhibit sufficient geometric diversity. This diversity can emerge either from the inherent physics of the system or by aggregating multiple distinct trajectories.
Consider the difference between an underdamped (oscillatory) system and an overdamped (non-oscillatory) system. Yuanyuan Wang et al. explore this in Physics from Video: Identifiability of Time-Invariant Second-Order ODEs under Minimal Trajectory Conditions. An oscillatory system naturally sweeps through the state space, visiting the same state levels multiple times with distinct velocities—a property the authors formalize as the level-set slope-coverage condition. This internal geometric diversity allows a single video clip to uniquely identify physical constants like damping and stiffness. In contrast, non-oscillatory systems lack this self-intersecting variety; they strictly require at least three independent trajectories with diverse initial conditions to resolve parameter ambiguities.
Irregularity as an Informational Probe
When moving from discrete trajectories to continuous distributions—such as mean-field particle systems governed by McKean-Vlasov models—the concept of an "initial state" shifts from a single point to an initial density distribution. Here, the geometric requirement manifests as a deconvolution condition based on smoothness.
Richard Nickl et al. demonstrate in Bayesian Nonparametric Inference in McKean-Vlasov models that the identifiability of hidden interaction forces depends fundamentally on the irregularity (roughness) of the starting density. A highly smooth, regular initial state allows diffusion to wash out and mask the underlying interaction effects during the system's evolution. Conversely, an irregular initial state acts as a high-frequency diagnostic probe. As the system evolves, the precise manner in which the dynamics smooth out this initial roughness uniquely betrays the underlying interaction potentials, making them identifiable.
Structural Pathways and Reachability
Beyond continuous geometry, the initial state also dictates whether information can flow through the system's underlying causal graph. In Identifiability in Continuous Lyapunov Models, Philipp Dettling et al. emphasize that identifying drift parameters requires the presence of sufficient "treks"—directed paths through common ancestors—between variables. Without these informational pathways being activated by the system's initialization, the observed covariance matrix lacks the mathematical constraints needed to distinguish between alternative configurations.
This structural view aligns with the findings of Alejandro F. Villaverde et al. in Structural Identifiability of Dynamic Systems Biology Models. By treating parameters as augmented states, the authors show that poorly specified initial conditions cause a catastrophic loss of reachability. If the initial states prevent the system from exploring certain dimensions, the underlying parameters remain unobservable, proving that the geometry of the starting state is the ultimate gatekeeper of system identification.
Key Foundational Work
Structural Identifiability of Dynamic Systems Biology Models (2016) — contributed 1 concept elements to this direction.
Identifiability of car-following dynamics (2021) — contributed 1 concept elements to this direction.