Boundary-Layer Stability and ISS

Define planner error $e=x-q^\star$. Differentiating and substituting the normalized filter gives

\[\epsilon\dot e=-e-\epsilon\dot q^\star.\]

This one equation contains both parts of the paper's claim: exponential convergence when the manifold is fixed and an $O(\epsilon)$ tracking tube when target motion carries it through state space.

Stationary reference

For stationary targets, $\dot q^\star=0$ and

\[\epsilon\dot e=-e.\]

With $V(e)=\tfrac12\lVert e\rVert^2$, its derivative is

\[\dot V=-\frac1\epsilon\lVert e\rVert^2 =-\frac{2}{\epsilon}V.\]

Proposition 1: uniform boundary-layer stability

For every fixed harmonic reference and every $\epsilon>0$, the equilibrium $x=q^\star$ is globally exponentially stable:

\[\lVert e(t)\rVert=e^{-t/\epsilon}\lVert e(0)\rVert.\]

The bound is uniform in the sheaf topology because $H$ has already been absorbed into the solved reference. The boundary-layer rate is $1/\epsilon$, not $\lambda_{\min}(H)/\epsilon$.

Measured stationary planner decay for three epsilon values overlaid with the predicted exponential rates

The fitted rates are $2.5$, $5.0$, and $10.0$ for $\epsilon=0.4$, $0.2$, and $0.1$ respectively. These are direct checks of the dissipation identity, not merely endpoint convergence.

Moving-reference ISS bound

Variation of constants gives

\[e(t)=e^{-t/\epsilon}e(0) -\int_0^t e^{-(t-s)/\epsilon}\dot q^\star(s)\,ds.\]

Taking norms yields

\[\boxed{ \lVert e(t)\rVert\le e^{-t/\epsilon}\lVert e(0)\rVert+ \epsilon\sup_{0\le s\le t}\lVert\dot q^\star(s)\rVert.}\]

The first term is the boundary-layer transient. The second is the asymptotic gain from target-induced reference velocity. Thus the planner is ISS with respect to $\dot q^\star$ and its uncompensated lag is first order in $\epsilon$.

This bound gives $\epsilon$ a physical interpretation. If the harmonic reference can move at most at speed $v_{max}$ and planner error must stay below $r$ after the transient, choose $\epsilon\le r/v_{max}$. The exact sampled update removes numerical stiffness, but it does not remove this continuous-time bandwidth tradeoff.

Tikhonov limit

On the fast time $\tau=t/\epsilon$, freeze $t$ and define $y=x-q^\star(t)$. The boundary layer is

\[\frac{dy}{d\tau}=-y,\]

which is globally exponentially stable uniformly in the frozen target state. If the reduced slow target/coordination dynamics satisfy the regularity and stability hypotheses of Tikhonov's theorem, then away from the initial layer

\[x(t)=q^\star(t)+O(\epsilon).\]

The theorem concerns the continuous system. It should not be confused with taking $\epsilon=0$ in software. TikhonovFilter requires a positive value; the algebraic direct-planner mode represents the exact zero limit.

From one agent to the fleet

Suppose local agent $i$ has tracking error $z_i$ and an ISS-Lyapunov function $V_i$ with respect to its delivered reference. Since local states are disjoint, define

\[V_{\mathrm{loc}}(z)=\sum_{i=1}^N V_i(z_i).\]

Directional differentiation is linear:

\[\dot V_{\mathrm{loc}}= \sum_{i=1}^N\dot V_i.\]

Summing the individual dissipation inequalities preserves positive definiteness and produces an ISS bound for the direct sum. Connecting that aggregate below the ISS planner forms the unidirectional cascade

\[p(t)\longrightarrow q^\star(t)\longrightarrow x(t) \longrightarrow(z_1,\ldots,z_N).\]

By the ISS cascade theorem, the complete system is ISS with respect to target motion and other admitted disturbances. When targets are stationary, the input vanishes; planner and local tracking errors converge to zero, giving the stationary-target GAS corollary.

What this does not prove

ISS of an unconstrained local controller does not automatically survive an arbitrary actuator projection, safety filter, or infeasible QP. Each inserted execution-layer map needs its own certificate or bounded-disturbance argument.