State space
State space is the set of all possible configurations that a system can occupy. It is not a physical space but a representational one: each point in the state space corresponds to a complete description of the system at a given moment, and the system's evolution over time traces a trajectory through this space. The concept is foundational to dynamical systems theory, cybernetics, control theory, and reinforcement learning — any discipline that treats systems as evolving rather than static.
The Structure of State Space
The state space of a system is defined by its state variables — the minimal set of quantities that, when specified, determine the system's future behavior given its dynamics. For a pendulum, the state variables are position and velocity; for a neural network, they are synaptic weights and activation values; for an economy, they might be prices, inventories, and expectations. The dimension of the state space equals the number of state variables, and this dimensionality has profound consequences for the system's behavior. High-dimensional state spaces permit complex adaptive dynamics — many interacting degrees of freedom that can produce emergence, self-organization, and phase transitions.
The geometry of the state space constrains what the system can do. A state space with multiple attractors is a system with multiple possible long-run behaviors; the initial condition determines which basin of attraction the trajectory falls into, and the boundaries between basins are the loci of sensitive dependence. A state space with a single attractor is a system that converges to a unique equilibrium regardless of initial conditions. The structure of the state space is therefore the structure of the system's possibility.
State Space and Information
In systems that learn or process information, the state space takes on an additional role: it is the space of possible representations. A predictive coding system does not merely occupy states; it occupies states that encode beliefs about the environment. The state space is stratified into regions that correspond to different hypotheses, and learning is the trajectory that moves the system from regions of high prediction error to regions of low prediction error. The state space is not just a mathematical convenience; it is the epistemic landscape within which the system navigates.
This reframes the problem of observability. A system is observable if its internal state can be reconstructed from its outputs — if the trajectory through the state space leaves a signature in the observable world that is sufficient to determine where in the space the system is. Observability is not guaranteed; many systems have hidden states that leave no trace in their outputs, and the reconstruction of these states requires either additional sensors or structural assumptions about the dynamics. The state space of a system and the observability of that space are two sides of the same epistemic coin.
State Space in Artificial Systems
In reinforcement learning, the state space is the foundation of the Markov decision process: the agent's policy is a mapping from states to actions, and the value function is a mapping from states to expected future reward. The RL agent's entire behavior is determined by how it partitions and navigates its state space. Yet the state space in RL is usually assumed to be fully known or fully observable — an assumption that is violated in nearly all real applications. Partially observable MDPs extend the framework by introducing a distinction between the true state space and the agent's belief space, a space of probability distributions over possible states. The agent no longer navigates the state space directly; it navigates a space of uncertainty about the state space.
This distinction is critical for understanding artificial systems. A language model does not have a state space in the sense of a physical system, but it has a representational space — the high-dimensional manifold of activation patterns that its parameters define. The geometry of this manifold determines what the model can represent, what it can generalize to, and what it will fail to capture. The study of this geometry is the study of the model's state space, and it is becoming central to the analysis of AI systems.
The state space is the most important concept in systems science because it is the place where ontology meets methodology. To specify a system's state space is to say what exists and what can change. To map that space — its attractors, its basins, its bifurcations — is to say what the system can do and what it will become. The mistake of classical reductionism was to look for the system's behavior in its components. The behavior is in the space. The components are merely the coordinates.