A Directional Relational Manifold is described by:
- a space
M; - a set of active directions
D(p)at each pointp; - an effective local dimension
dimD(p) = |D(p)|or a soft relaxation by gates; - a relational metric
g_pover admissible directions; - admissible curves whose velocities lie in
span(D(p)).
In this implementation, p is represented by a latent state vector z.
The model learns directions:
V_i(z) in T_z M
a_i(z) in [0, 1]
dimD(z) = sum_i a_i(z)
The directions may be non-orthogonal. Gates provide a soft active set instead of a fixed fundamental dimension.
The learned metric is:
G(z) = diag(softplus(d(z)) + eps) + U(z)U(z)^T
For a velocity v, metric energy is:
g_z(v, v) = v^T G(z) v
The low-rank term gives the metric a learned coupling structure between state coordinates and directions.
A rollout is admissible when:
dz_t in span(D(z_t))
The implementation enforces this by constructing:
dz_t = sum_i a_i(z_t) c_i(z_t, e_t) V_i(z_t)
The discrete action proxy is:
A(z_0:T) = sum_t dt * g_z_t(dz_t, dz_t)
Training can penalize mean action. This is a variational bias, not an exact geodesic solver.
The MVP uses an Euler update:
z_{t+1} = z_t + dt * dz_t
When bounded_state is enabled, norm clipping and tanh projection keep the coordinate state in a compact region.
A future version could define relational metrics through pullbacks from decoder distributions or Fisher information. That would connect the geometry more tightly to emitted language distributions.
Toroidal convergence is not assumed. It is only a possible hypothesis under strong conditions such as boundedness, recurrence, and structural stability. The optional toroidal representation is a coordinate utility, not evidence of toroidal topology.