Library.Reference:6.3 - A faster command can defeat an unchanged controller
Consider a constructed sampled controller with plant x[k+1] = x[k] + u[k] and tracking law u[k] = (r[j] - x[k])/2. The planner’s command r[j] is held for N fast steps. The state is fully observed; this model has no disturbance, computation delay or unmodeled saturation. Commands satisfy |r[j]| ≤ 8.9; consecutive command changes have magnitude at most δ. Required sampled limits are |x[k]| ≤ 10 and |u[k]| ≤ 0.6.
Planning and tracking are two functional contributions that could run on one processor. A separate system composition could include a controller, plant and their constituents. The functional grouping alone does not establish two physical Systems or two system-composition levels. The current question is whether the command and tracking contributions are compatible under the sampled limits.
Let e[j] = x[jN] - r[j] immediately after a command is selected. During the held-command interval the error halves at each fast step. At the next command change,
e[j+1] = 2^(-N) e[j] - (r[j+1] - r[j]).
If the initial error satisfies |e[0]| ≤ E, the interval |e| ≤ E at command changes is invariant for E = δ/(1 - 2^(-N)): the next error is at most 2^(-N) E + δ = E. Within each interval the error decreases. Hence |u| ≤ E/2, and |x| ≤ |r| + E is a sufficient state bound. This is the consequential join: the planner’s jump and update conditions determine the tracking demand.
With N = 4 and δ = 1, E = 16/15. The input magnitude is at most 8/15 < 0.6, and the sufficient state bound is 8.9 + 16/15 = 299/30 < 10. Under the stated initial and model conditions, every command sequence within this interface respects the sampled limits. Holding the current command is allowed by those constraints; the calculation does not establish attainment of every further planner objective.
Now require commands every two fast steps while retaining permitted unit jumps. The plant, processor and two functional contributions are unchanged. The error bound becomes E = 4/3, so the earlier sufficient input argument fails. There is also an actual violating trajectory: take r[0] = 0 and x[0] = 16/15, admitted by the earlier design. After two steps, x[2] = 4/15. The permitted command r[1] = -1 then gives u[2] = -19/30, whose magnitude exceeds 0.6. The failure is in the changed temporal and command interface, not in the number of component boxes.
One repair keeps N = 2 and the command-magnitude bound 8.9, but reduces the jump bound to δ = 0.9. Require initial conditions |e[0]| ≤ 1.2 and |x[0]| ≤ 10. The error envelope is invariant because 1.2/4 + 0.9 = 1.2, so |u| ≤ 0.6. Check the state separately: x[k+1] = (x[k] + r[j])/2 is a convex combination. From |x[k]| ≤ 10 and |r[j]| ≤ 8.9, it follows that |x[k+1]| ≤ 9.45 ≤ 10. Thus the state set is invariant as well.
The coarser bound |r| + E would have suggested reducing the command magnitude to 8.8. The separate state invariant shows that this extra restriction is unnecessary in the stipulated model. The revised interface still sacrifices the original unit-jump demand, and a transition from an operating controller must establish both new initial conditions. Accept this candidate only if the intended task permits that smaller jump. Retaining N = 4 instead sacrifices the faster response. A timed command segment, another tracking law or stronger actuation changes other premises and requires another derivation.
CAD.1–CAD.4 in Control Architecture Design develop the corresponding architecture, interface, feasibility and co-design questions. This sampled calculation leaves real-plant adequacy, continuous-time constraints, shared-processor scheduling and the planner’s further objectives to their required constructions. Keeping only two boxes loses the recurrence and cannot decide this change; retaining the recurrence, command conditions and initial state is enough for the demonstrated sampled-limit comparison. Processor detail becomes relevant if computation or delay changes that premise.