Lecture 6 – Nonlinear controllability Material Nonlinear System

ẋ = f (x, u)
y = h(x, u)
Lecture slides
Handout from Nonlinear Control Theory, Torkel Glad (Linköping)
Nonlinear Controllability Important special affine case:
Handout about Inverse function theorem by Hörmander
ẋ = f (x) + g(x)u
y = h(x)

f : drift term
g : input term(s)

What you will learn today (spoiler alert) What you will learn today What we will not do
Local Controllability:

New mathematical concepts and language A nonlinear system is controllable if the linearized system is
Local Observability. Depends on x0 and u.
controllable.
ẋ = f (x) + g(x)u is “accessible” iff yj = hj (x)
Manifolds, charts (M, „(x)) O = spanLX1 . . . LXk hj (x)
dim (f, g, [f, g], [f, [f, g]], . . .) = n
n
ÿ ˆ dO = span (dH | H œ O)
Vector fields ai (x)
i=1
ˆxi
Fundamental Parking Theorem The system is locally observable if
n
ÿ ˆf
Lie-derivative LX (f ) = ai (x)
ˆxi dim (dO) = n
i=1
ˆg ˆf
Lie-bracket [f, g] = f≠ g Duality between observability and controllability
ˆx ˆx

And not this and not this Basic Result: Linearization at (x0 , u0 )
Nonlinear Kalman Decomposition
Can find coordinates (x1 , x2 , x3 , x4 ) so that ẋ = f (x) + g(x)u, x(0) = x0
ẋ1 = f 1 (x1 , x3 ) + g(x1 , x3 )u
Differential Flatness Theorem Suppose f (x0 ) + g(x0 )u0 = 0. If the linearization
ẋ2 = f 2 (x1 , x2 , x3 , x4 ) + g(x1 , x2 , x3 , x4 )u
3 Zero Dynamics
ẋ3 = f (x3 ) ż = Az + Bv
ẋ4 = f 4 (x1 , x3 ) Nonlinear Minimum Phase ˆf ˆg
A = (x0 ) + (x0 )u0
y = h(x1 , x3 ) Disturbance Decoupling
ˆx ˆx
B = g(x0 )
f h(x0 ) ”= 0
Relative Degree Smallest r such that Lg Lr≠1 Normal Forms
is controllable, then for all T > 0, ‘ > 0 the set
Exact Linearization by Feedback Stabilization
ẋ = f (x) + g(x)u XT,‘ = {x(T ); |u ≠ u0 | < ‘}
u = –(x) + —(x)v and z = Z(x) =∆ ż = Az + Bv
contains a neighborhood of x0 . (Proof: Nice exercise in using the
System is feedback linearizable if one can find y = h(x) so the inverse function theorem)
system has relative degree n. Can be checked with Lie-brackets
 Manifolds Example - Pendulum Rigid Bodies
What are natural mathematical models for state spaces?

Natural State Space

Y Z
R = [r1 r2 r3 \ œ SO(3)
ie RRT = I and det(R) = 1
◊¨ = sin(◊) + u Ṙ = ≠R ◊ w … Ṙ = ≠RS(w)
Y Z
Natural state space: R ◊ S 1 = cylinder _ 0
_
_
≠w3 w2 _ _
_
Piece together “bent” pieces of Rn . S(w) = _
_
_
_ w3 0 ≠w1 _
_
_
_
[ \
≠w2 w1 0
Same local properties as Rn . Different globally S 1 = unit circle

Rolling Penny Rolling Penny Definition of Manifold

Yes it can. But it is not obvious.

ẋ = u1 cos(◊) The penny has non-holonomic constraints a(z)ż = 0
ẏ = u1 sin(◊) Y Z
Ï̇ = u1 Y Z__ ẋ _
_
_
_ ẏ _
sin ◊ ≠ cos ◊ 0 0_ _
◊˙ = u2
_ __ _
_=0
_
_ _ _
_ _
_
[ \_
cos ◊ sin ◊ ≠1 0 __
_Ï̇_ _
[ ˙_ \
◊
A C Π(=smooth) manifold is a topological space M together with an
The linearization is not controllable (check)
atlas {U– , Ï– } of pairwise C Œ -compatible coordinate charts that
Can the penny be moved sideways in small time (keeping the head Can not be written as holonomic constraints: h(z) = 0 =∆ hz ż = 0. cover M .
up)?
Topological space M ?
Atlas {U– , Ï– } ?
Pairwise C Π-compatible coordinate charts ?

Topology Compatible Coordinate Charts Example: Cylinder

A topology on a set M is a collection T of subsets of M .

O is called “open” if O œ T .
The collection T must be such that

ÿ, M œ T
O1 , O2 œ T =∆ O1 fl O2 œ T
{Oi } œ T =∆ fiOi œ T
 ¶ Ï≠1 smooth on U fl V = (x2 ”= 0, z2 ”= 0)
Compatible:  ¶ Ï≠1 (x) œ C Œ
f is called ”smooth” if f ( ≠1 (x)) œ C Œ , ’ z = Â(Ï≠1 (x)) is given by (z1 , z2 ) = (x1 , 4/x2 )
The cylinder is a smooth manifold
Note: f ¶ Ï≠1 (x) = f ¶  ≠1 ¶  ¶ Ï≠1 œ C Œ
Independent on coordinate charts.
 Examples Global Differences to Rn - Example Manifolds defined by equation systems

Many manifolds are defined implicitly by equations systems

f1 (x1 , . . . , xn ) = 0
..
.
fk (x1 , . . . , xn ) = 0

When does this describe a (smooth) n ≠ k -dimensional manifold?

Any smooth velocity field v on S 2 must have a point where v(x) = 0

“You can’t comb the hair of a tennis ball”

Differentials Differentials Inverse Function Theorem

Y ˆf1 Z Theorem Let X be open in U and f œ C 1 (X, V ), f (x0 ) = y0 . For

f : A æ B is called differentiable at x œ A iff there is a continuous existence of g œ C 1 (Y, U ) where Y is a neighborhood of y0 so
ˆf1
_
_ ˆx1 ˆx2 . . ._
_
_
_ _
_
linear map DFx (h) : A æ B such that _ ˆf2 .. _
DF = _
_
_
_ ˆx1
. _
_
_
_ a) f ¶ g = identity near y0
_
_ . _
_
[ .. \
Îf (x + h) ≠ f (x) ≠ DFx (h)Î æ 0, hæ0 b) g ¶ f = identity near x0
c) a) and b)
DFx = differential (Jacobian) Definition Rank of f at x := rank(DFx ).
If f smooth then it is necessary and sufficient that there is a linear map A such that
Y ˆf1 ˆf1 Z respectively
_
_ . . ._
_
_ ˆx1
_
_ ˆf2
ˆx2
..
_
_
_ Rank (DFx0 ) = k =∆ Rank (DFx ) Ø k
DF = _
_
_ . _
_
_ a’) f Õ (x0 )A = IV
_
_ ˆx1 _
_
_ . _
[ .. \ for all x close to x0 . b’) Af Õ (x0 ) = IU
Proof: Dk (x) = k ◊ k submatrix of DFx with c’) a’) and b’)
det(Dk (x0 )) ”= 0 =∆ det(Dk (x)) ”= 0, for x close to x0 .
Condition c’ implies that g is uniquely determined near y0 .

Implicit Function Theorem Example

h(x, y) = 0
Proof idea: To solve y = f (x) use ˆh
full rank =∆ x = x(y) uniquely
ˆx
xk = xk≠1 + f Õ (x0 )≠1 (y ≠ f (xk≠1 ))
q
Prove (xk ≠ xk≠1 ) converges for y near y0 .
h(x, y) = x2 + y 2 ≠ 1, hÕx = 2x
See handout.

So x = x(y) uniquely except near (0, ±1).

 
In fact x = 1 ≠ y 2 , x0 > 0 and x = ≠ 1 ≠ y 2 , x0 < 0.
 Discussion Functions Between Manifolds
Note
Implicit F. T. =∆ Inverse F. T. c). f (x, y) = 0
ˆx
h(x, y) = y ≠ f (x); hÕx = fxÕ =∆ x = x(y) uniquely fxÕ + fyÕ = 0
ˆy
ˆx
= ≠(fxÕ )≠1 fyÕ
ˆy
Inverse F. T. c) =∆ Implicit F. T.

f (x, y) = (h(x, y), y) Try it yourself example

Y Z
Õ
_hx
_ hÕy _ x31 ≠ ex2 + x33 ≠ 1 = 0
fÕ = _
[
_
_
\ full rank
0 I x21 + x2 ≠ x23 = 0
So (x, y) locally determined by (h, y) = (0, y) Definition
Are x1 , x2 smooth functions of x3 around (1, 0, 1)?
What is ˆx 1
ˆx3 at that point? f œ CŒ ≈∆  ¶ f ¶ Ï≠1 œ C Œ , ’Â, Ï
=∆ x = x(y) uniquely locally

Manifolds by equation systems Tangent Vectors - different definitions Our Definition

Derivative operator X(f ) : (f : M ‘æ R) ‘æ R

f1 (x1 , . . . , xn ) = 0 X(–f + —g) = –X(f ) + —X(g)

.. X(f g) = f X(g) + gX(f )
.
fk (x1 , . . . , xn ) = 0 Example: Take any coordinate chart (U, Ï) with coordinates x. Then
determines an n ≠ k dimensional manifold near x̄ if Define it only for manifolds embedded in Rn : n
ÿ ˆ
Y ˆf1 ˆf1 Z Xa = –i
Ï(t) ≠ Ï(0) ˆxi
_
_
_ ˆx1 ˆx2
. . ._
_
_ ẋ = lim i=1
_
_ .. _
_ tæ0 t
_
_ ˆf2
_ . _ has full rank (= k) at x̄
_
_ is a tangent vector, where
_ ˆx
_ .1 _
_ Velocity vectors in Rn .
_
[ . _
\
. Coordinate free version. Tangent vectors at x ¡ ”equivalance n
ÿ ˆf
classes of curves with Ï(0) = x”, we define Ï(t) ≥ Â(t) when Xa (f ) = –i (a)
i=1
ˆxi
(”Gradients Òf1 , . . . , Òfk are linearly independent”) Ï(t) ≠ Â(t)
Ï(0) = Â(0) = x and lim = 0 in some chart
tæ0 t

Coordinate Change Push Forward Operator

Y Z
Y Z_
_ –1 _
_ . _
ˆ \_
_
X = [ ˆxˆ 1 ... ˆxn
_
_
_ .. _
_
_
_
Different notation [ _
_ \
–n
LX (f ) = X(f ) Lie-derivative = fishermans derivative Change coordinates — = ˆx
ˆz
– or
Y Z Y Z ˆz
[fú X](g) := X(g ¶ f )
Examples [ ˆ ˆ \
ˆx1 . . . ˆxn = [ ˆzˆ 1 ... ˆ \
ˆzn
ˆ ˆ ˆ ˆ ˆx
; ; z + sin(◊)
ˆ◊ ˆz ˆ◊ ˆz
Example

z1 = x1
z2 = x1 + x2
ˆ ˆz1 ˆ ˆz2 ˆ ˆ ˆ
= + = +
ˆx1 ˆx1 ˆz1 ˆx1 ˆz2 ˆz1 ˆz2

Note that x1 = z1 does not imply ˆx

ˆ
1
= ˆzˆ 1
 (Smooth) Vector Fields Integral Curve Transformation Group, Flow
‡(t) is an integral curve to X if in local coordinates
Y Z
_
_ ‡1 (t) _
_
_
_ . _
‡(t) = _
_ . _ _
_
_ . \
_
[
_
_
‡n (t)
ˆ
(g(‡(t)) = X(‡(t))(g)
ˆt
ÿ ˆg d‡i ÿ ˆg
Assigns a tangent vector to each point: p ‘æ Xp = Xi (‡(t))
ˆxi dt ˆxi X t (p) = solution to ẋ = X(x), x(0) = p
n
ÿ ˆ i.e.
X= Xi (p) Xt is smooth. X 0 = id
ˆxi ‡˙ 1 = X1 (‡(t)) g(X h (p)) ≠ g(p)
i=1 n
ÿ ˆg
Xi (p) smooth functions of p. .. LX (g) = X(g) = Xi = lim
. ˆxi hæ0 h
Y Z i=1
_ X1 (x1 , . . . , xn ) _
_
_ _
_ ‡˙ n = Xn (‡(t)) L–X+—Y = –LX + —LY , –, — œ R
_ .. _
Alternativ notation: _
X≥_
_ .
_
_
_
_ _
[ \
Xn (x1 , . . . , xn ) ẋ = f (x, u) f : M ◊ U ‘æ T M
A set of ODEs

Example Main new object: Lie Bracket of vector fields Why is it interesting?

ẋ = f (x) + g1 (x)u1 + g2 (x)u2 + . . .

ẋ = f (x) + g(x)u Consider two vector fields ẋ = f (x) and ẋ = g(x)
y = h(x)

Lie-bracket. New vector field

ˆh ˆh ˆg ˆf
ẏ = ẋ = (f + gu) = Lf +gu h [f, g] = f≠ g
ˆx ˆx ˆx ˆx
= Lf h + uLg h
y (k) = (Lf +gu )k h Roughly we have:
If the Liebracket ”tree” has full rank, then the system is ”controllable”.

Example Example Example

ẋ1 = u1 Hence at x = 0 we have

ẋ2 = u2
Y Z Y Z Y Z ẋ = f (x) + g(x)u
_
_ 1_ _ 0_ _ 0 _
_ __ _
_ _ _ _
_ _
_ y = h(x)
_0_
g1 = _ 1_ 0 _
ẋ3 = x1 u2 ± x2 u1 _
_ _ , g2 = _
_ _
_ _
_, [g1 , g2 ] = _
_
_ _
_
[ _
\ _
[ _ \ _
[ _
\
0 0 1 ≠ ±1
Y Z Y Z
_
_ 1 _
_ _
_ 0_ With the minus-sign the three vector fields span R3 , and we have
_ _ _ __
This means g1 = _
_
_
_ 0 _
_
_
_ and g2 = _
_
_
_ 1_
_
_ controllability. ˆh ˆh
[ \ [ _\ ẏ = ẋ = (f + gu) = Lf +gu h
±x2 x1 ˆx ˆx
With the plus-sign the system is not controllable, in fact it can be seen
Y ZY Z Y ZY Z = Lf h + uLg h
_0 0 0_ _ 1 _ _0 0 0_ _0_ that x21 + x22 ≠ 2x3 is an invariant.
_
_ _
_
0 0 0_
_
_ _
_ _
_ 0 _
_ _
___ _
_ y (k) = (Lf +gu )k h
_0 0 0_
[g1 , g2 ] = _
_
_
_
_
_ _ ≠_
_ _ _1_
_ _
_
[ __
\ _
[
_
_
\ [
_ _
_
\_
_
[ _
_
\
1 0 0 ±x2 0 ±1 0 x1
 Lie-Brackets Another example Lie-Brackets
Why are Lie-brackets so fundamental?
ẋ = g1 u1 + g2 u2
Y
Y Z _
_ (1, 0) t œ [0, h)
[X, Y ]p (f ) = Xp (Y (f )) ≠ Yp (X(f )) _cos „_ ˆ ˆ _
] (0, 1)
X =
_
_ _ ≥ cos „
_ +r t œ [h, 2h)
[
r \ (u1 (t), u2 (t)) =
Y Z
ˆr ˆ„ _
_ (≠1, 0) t œ [2h, 3h)
_
[
Y Z Y Z _≥r ˆ + ˆ
_r _
_ (0, ≠1) t œ [3h, 4h)
_ X1 _
_ _ Y1 _
_ Y = _
[ _
_ _
_ .. _
_ _ _
_ .. _
_ 1\ ˆr ˆ„ x(4h) = x0 + h2 [g1 , g2 ] + O(h3 )
X≥_
_ . _
_ _;
_ Y ≥_
_ . _
_ _
_ Y ZY Z Y ZY Z
_1 0_ _cos „_ 0 ≠ sin „_
_
[ _ _
[ _ _ _r _
\
Xn
\
Yn [X, Y ] = _ __
_ _ _≠_
_ _
_ _
__ _ __
[
0 0\ [ r \ [1 0 \ [1\ Trotters Product Formula
Y Z A 
ˆY ˆX  t Bn
_cos „ ≠ sin „_
t t
[X, Y ] = X≠ Y =
_
_ _ ≥ (cos „ ≠ sin „) ˆ ≠ r ˆ
_
t
ˆx ˆx [
≠r \
ˆr ˆ„
t
[X,Y ] = lim ≠Y
n
≠X
n
Y
n
X
n
næŒ

Proof sketch 3 4n
tf 1 tf 2
1+ +o æ etf
n n

Some Lie-Bracket Formulas Vector Fields, Summary

A vector field X is associated with
a) A system of differential equations
[f X, gY ] = f g[X, Y ] + f X(g)Y ≠ gY (f )X
[X, Y ] = ≠[Y, X] dx a æ b solution to differential equations
= X(x)
dt
[X1 , [X2 , X3 ]] + [X2 , [X3 , X1 ]] + [X3 , [X1 , X2 ]] = 0 b æ c direct
1 b) A flow t :M æ
‘ M, t œ [t0 , t1 ], where ‡(t) = t (x) is the
LX Y = [X, Y ] = lim [Xú≠h Y ≠ Y ] solution to c æ d direct
hæ0 h d‡
= X(‡), ‡(0) = x d æ e proposition
dt
c) A directional derivative e æ a direct
Œ
ÿ hn h2 d -
Xú≠h Y = adn
XY = Y + h[X, Y ] + [X, [X, Y ]] . . . Xx f = f ( t (x))-
-
n=0
n! 2 dt t=0
d) A ”derivation” of the algebra C Œ (M ).
related to
e) A partial differential operator
1
eA eB = eC ; C = A + B + [A, B] + . . . ÿ ˆ
2 X= Xj
ˆxj

Park Your Car Using Lie-Brackets! Parking cont’d

Y Z
_≠ sin(„ + ◊)_
_
_ _
_
_
_ cos(„ + ◊) _
_
[Steer, Drive] = . . . = _
_
_
_ := Wriggle
_
_
_ cos(◊)
_
_
_
_
_
[ \
0

(x, y) : position Y Z Y Z
: direction of car _
_0_ _cos(„ + ◊)_
„ _ _
_ _
_ _
_
_0_
_ _ _ sin(„ + ◊) _
_ _
◊ : direction of wheels Steer = _
_ _
_ _,
_ Drive = _
_
_
_
_
_
_
_0_
_ _
_
_
_
_ sin(◊) _
_
_
(x, y, „, ◊) œ R2 ◊ S 1 ◊ [◊min , ◊max ] [ \
1
[
0
\
An easy calculation (exercise) shows that
Linear Systems
Y Z
_
_≠ sin(„ + ◊)_
_
_ _
_ cos(„ + ◊) _
_ _
[Wriggle, Drive] = __
_ _ =: Slide
_
_ ẋ = Ax + Bu = f (x) + g(x)u
_
_
_ 0 _
_
_
[ \
0
Fundamental Parking Theorem
For ◊ = 0 this takes you sideways: You can get out of any parking lot that is larger than the car. Use the
following control: Wriggle, Drive, –Wriggle (this requires a cool head),
–Drive (repeat).
Proof: Trotters Product Formula

[f, g] = [Ax, B] = 0 ≠ AB
[g, [f, g]] = 0
[f, [f, g]] = [Ax, ≠AB] = A2 B
..
.
Slidet (x, y, „, 0) = (x ≠ t sin(„), x + t cos(„), „, 0)
Adkf g = [f, [f, . . . , [f, g]]] = (≠1)k Ak B
¸ ˚˙ ˝
k Lie-brackets

Controllability Theorems Accessibility theorem Reading Assignment

C = smallest Lie subalg. containing {f, g1 , . . . , gm }

q
ẋ = f (x) + i gi (x)ui
For a precis formulation, and more about ”controllability” vs
Let A(x0 ) be the reachable set from x0 , i.e. all points that can be
”accessability” see
reached from x0 using a suitable control u
Accessibility The system has the accessibility property at x if A(x) T. Glad, Nonlinear Control Theory, Chapter 8, pp 73-81
has nonempty interior
Theorem If for all x0 the Lie-bracket tree contains n linearly
independent elements, then the system is has the accessbility property

dim C = n =∆ can reach open set

If f = 0, (or more generally f (x, u) is ”symmetric”, see Glad) then the

system is controllable: A(x0 ) = Rn