Stability Theory

When a system has control, then comes the questions of whether we could make it stable under the control.

Problem Setup

A control affine system is this

˙x=f(x)+igi(x)ˉui=f(x)+g(x)ˉu

Interpretation:

  • This control system has multiple vector field, a drift field f by default, and a few “Gain” field that you can control gi.
  • You can modulate those “Gain” fields and leverage the drift fields to navigate the space.
  • Note in the degenerate case, if f=0 and m=1, then ˙x=ˉug(x), this system is driftless, and the control signal can only modulate speed / time, cannot change the track.

The equilibrium stablization problem is: For a controlled equilibrium 0, f(0)+g(0)u=0, find a feedback function u:RnRm, s.t. x=0 is a globally asymptotic stable equilibrium.

Regular Feedback: regular feedback is a function ˉu on state space Rn, 1) Locally Lipschitz on Rn0 2) continuous at 0

  • Regular feedback is continuous.

Condition for Existence

Main theorem is Arstein Sontag theorem

Arstein Sontag theorem: For ˙x=f(x)+g(x)ˉu, f(0)=0, there exists a global regular stablizer if and only if, there exist a C1 contrrolled lyapnov function (CLF), which satisfied small controlled property.

  • Note this theorem states only continuous feedback. There are scenarios that requires in-continuous or more complex feedback like those with internal dynamics.

Controlled Lyapnov Function: A function V:RnR, that is pos. def. at 0 and radially unbounded. Also satisfies the equivalent conditions

  • When x0, and when Lg1V=Lg2V=LgiV=0, then LfV0
  • x0, then there uRm s.t. LfV+iuiLgiV<0
    • Or x0, infuRm[LfV+iuiLgiV]<0

Interpretation:

  • This C.L.F.’s property and existance is mandated by the converse Lyapnov theorems.
  • There exists at least a control vector everywhere to decrease the C.L.F.
  • When all the control signal cannot decrease C.L.F. (geometrically, the gradient of C.L.F. is orthogonal to all the vector fields gi), then the drift field can still decrease C.L.F.

Small Control Property For a C.L.F., RR, ϵ>0,δ>0, such that, xBδ(0)0, there uBϵ(0), s.t. LfV+iuiLgiV<0.

Interpretation: When the state deviation from equilibrium is small, you can use a small control to make it converge (decrease the C.L.F.). (the Columb repulsion is like a counter example.)

General Results

Brockett’s Necessary Condition: For system ˙x=f(x,u), f is locally Lipschitz w.r.t. x and u. If it has a continuous stablizer ˉu(x), ˉu(0)=0. then ϵ>0,δ>0, s.t. Bnδ(0)f(Bnϵ(0)×Bmϵ(0)).

Interpretation:

  • Any neighborhood around 0 in the tangent space, could be covered by the image of the function f(x,u) on a small neighborhood of control and state.
  • In a small neighborhood around 0, small control signal and the drift can create vector ˙x to any direction / small amplitude! (This system is fully steerable)
    • In a sense, the velocity created is not limited to a subspace.

Corrolary for driftless affine system: If f=0 and ˙x=iˉuigi(x), then if the gi vectors are rank deficient rank([g1(0),g2(0),,gi(0)])<n then the 0 cannot be stablized continuously.

Controllability for LTI system: For linear control system, ˙x=Ax+Bu, then the controllabilty is determined by the Kalman rank condition: rankC=n, C=[B,AB,A2B,,An1B]

Interpretation:

  • This is a corrolary of Brockett theorem for LTI.
  • The controllability matrix characterize the interaction between the control gain and the drift.

Passivity

Intuition: The system doesn’t generate energy, the energy increase in it is less than or equal to the energy input.

  • This property is inspired by the energy property in a electrical / mechanical system.

Passivity Based Control

MIT Dynamic System and Control 2011