From encoding to dynamics
Everything you decoded in the earlier volumes leaned on an encoding picture: a neuron in motor cortex has a cosine tuning curve for movement direction, a population vector sums those preferences, and a Kalman filter reads the intended velocity back out. It works — but it quietly assumes each cell is a labelled sensor for an external variable. The research frontier rejects that framing. It asks instead: what if the cortex is not representing movement so much as generating it, as the output of an internal dynamical system whose job is to produce the right time-varying pattern of muscle commands?
This is not a semantic quibble. Single-cortical responses are notoriously mixed and heterogeneous: a neuron may co-vary with direction, speed, and time in ways that defy a clean tuning label. The dynamical-systems view treats that heterogeneity as expected — the apparent tuning of any one cell is an epiphenomenal shadow of a lower-dimensional collective process. What is lawful lives not in the neuron but in the population state and how it moves.
The neural manifold
The empirical anchor for all of this is the neural manifold. Record hundreds of neurons and you might expect the population to wander freely in a hundred-dimensional space; instead, during structured behavior the activity is confined to a curved surface of far lower dimensionality — often a handful to a few dozen dimensions. The population state is a point on that surface; behavior is a trajectory traced across it over time.
A working measure of effective dimensionality — the participation ratio of the covariance eigenvalues λᵢ (from PCA). When a few components dominate the variance, d_PR is small: the manifold is low-dimensional.
Dynamics as computation
If the state lives on a manifold, the deep question is how it moves. The dynamical-systems hypothesis says the population obeys an autonomous (or input-driven) flow: from any state, the next state is largely determined, the way a pendulum's future follows from its position and momentum. Motor preparation, on this view, sets an initial condition in an optimal subspace, and movement is the system releasing from that condition and rolling downhill.
The core object of the whole track: a latent state x of dimension k — far smaller than the neuron count N — evolving under a flow F, optionally driven by inputs u. Everything ahead is a way to learn F and x from spikes.
This reframing pays off empirically. Motor-cortical trajectories show striking rotational structure — the state cycles through phase space like an oscillator — and states are arranged to avoid trajectory tangling, keeping the flow field smooth and noise-robust the way a well-designed dynamical machine would. These are population-level signatures with no natural description in the single-neuron tuning language.
Why a BCI should care
Beyond elegance, latent dynamics buy the engineer concrete things. A learned dynamical model is a prior: it tells the decoder which states and transitions are plausible, denoising moment-to-moment estimates far better than a memoryless readout. The latent state is also a lower-dimensional, more stable decode target than raw firing rates — the substrate that latent-space stabilized decoders exploit to survive drift for months. And a dynamical model that generalizes is the natural backbone for the transfer and foundation-model ambitions of the neighbouring tracks.
So this track is not a detour into applied math. It is the field's attempt to answer its deepest scientific question — how does neural activity compute? — with tools rigorous enough that the answer also improves the interface. The rest of the guides build that toolkit from linear and probabilistic foundations, through deep nonlinear models, to the hard question of whether any of it is truly identifiable.