Hankel rank and minimal linear realizations

Read side by side

Field-valued semantics admit linear state representations that can be much smaller than a discrete observer storing exact future behavior.

For a weighted series f : Sigma* -> K, define its residual after prefix u by

f_u(v) = f(uv).

A discrete future-equivalence observer identifies prefixes only when the full residual functions are equal. A linear realization instead stores a vector in the span of the residual functions. Over a field, the dimension of this span is the rank of the Hankel matrix H[u,v] = f(uv) and equals the state dimension of a minimal weighted finite automaton. This is the classical weighted-automata and rational-series correspondence.

A polyphase example

Take the polyphase route packet for the two-tap filter h=(1,1), polyphase order N, and evaluate the coarse-lag polynomial variable at z=2. The resulting route matrix is

M_N = I + S, with S^N = 2 I.

For the scalar downstream query

f_N(n) = e_0^T M_N^n e_0,

the N x N Hankel block has full rank N. Thus the existing N-phase linear route state is already minimal for this query.

For N=2,

M = [[1,2],[1,1]]

and

f(n) = 1,1,3,7,17,41,...,

which satisfies

f(n+2)=2 f(n+1)+f(n).

All discrete residual functions k -> f(n+k) are distinct, so a deterministic Moore observer storing exact future behavior requires infinitely many states. The linear/Hankel representation requires dimension two.

This example illustrates why weighted automata and linear state-space models can compress numerical series more effectively than deterministic state equality.

Architectural consequence

The continuation quotient must be typed by the observation algebra:

API boundary

The polyphase route packet supplies the finite linear realization used here. The recurrence subsystem supplies companion-state coordinates. A generic weighted-automaton constructor is outside these APIs. Minimality depends on the observation algebra, as described in carrier-change obstructions.