Massachusetts Institute of Technology

Department of Electrical Engineering and Computer Science

6.243j (Fall 2003): DYNAMICS OF NONLINEAR SYSTEMS
by A. Megretski

Take-Home Test 1 Solutions1

Problem T1.1
Find all values of µ ≤ R for which the function V : R2 ∞≈ R, defined by
�� �

x̄1
V = max{|x̄1 |, |x¯2 |}
x¯2
is monotonically non-increasing along solutions of the ODE
�
ẋ1 (t) = µx1 (t) + sin(x2 (t)),
ẋ2 (t) = µx2 (t) − sin(x1 (t)).

Answer: µ ∀ −1.
Proof For µ ∀ −1, x1 ∈= 0 we have
1d 2
x = µx21 + x1 sin(x2 )
2 dt 1
< −|x1 |(|x1 | − |x2 |),
and hence |x1 | is strictly monotonically decreasing when x ∈= 0 and |x1 | √ |x2 |. Similarly,
|x2 | is strictly monotonically decreasing when x ∈= 0 and |x2 | √ |x1 |. Hence, when µ ∀ −1,
V (x) is strictly monotonically decreasing along non-equilibrium trajectories of the system.
For µ > 1, x1 (0) = r, x2 (0) = r, where r > 0 is sufficiently small we have
ẋ1 (0) = µr − sin(r) > 0,
hence
V (x(t)) √ x1 (t) > r = V (x(0))
when t > 0 is small enough, which proves that V is not monotonically decreasing.

1
Version of October 20, 2003
 2

Problem T1.2
Find all values of r ≤ R for which differential inclusion of the form

ẋ(t) ≤ �(x(t)), x(0) = x̄0 ,

2
where � : R2 ∞≈ 2R is defined by

�(¯
x) = {f (¯
x/|x|)}
¯ for x¯ =
∈ 0,

�(0) = {f (y) : y = [y1 ; y2 ] ≤ R2 , |y1 | + |y2 | ∀ r},

has a solution x : [0, →) ∞≈ R2 for every continuous function f : R2 ∞≈ R2
2
and for every � initial condition x¯0 ≤ R .
Answer: r √ 2.
�
Proof First, let us show�that existence of solutions is not guaranteed when r < 2. Let
τ > 0 be such that 2τ < 2 − r. Define
�� �
� � �
x1 0.5�2(1 − τ) − x1
f = .
x2 0.5 2(1 − τ) − x2

Let us show that, for this f , the differential inclusion ẋ(t) =≤ �(x(t)) will have no solutions
x : [0, →) ∞≈ R2 with x(0) = 0. Indeed, since

∈ 0, x ≤ R2 ,
x� f (x/|x|) < 0 � x =

|x(t)| is strictly monotonically decreasing when x(t) ∈= 0. Therefore � x(0) = 0 implies

x(t) = 0 for t √ 0. Hence ẋ(t) = 0 ≤ �(0). However, for r < 2, and for this particular
selection of f (·), zero is not an element of �(0). The contradiction shows that no solution
with x(0) = 0 exists. �
To prove existence of solutions for r √ 2, one is tempted to use the existence theorem
relying on convexity and semicontinuity of �(·). However, these assumptions are not
necessarily satisfied in this case, since the set �(0) does not have to be convex. Instead,
note that, by the continuity of f , existence of a solution x : [t0 , t0 + |x̄0 |/M ) ∞≈ R2 with
x(t0 ) = x¯0 is guaranteed for all x¯0 ∈= 0. Hence, it is sufficient to show that a solution
x0 : [0, →) ∞≈ R2 with x0 (0) = 0 exists.
To do this, consider two separate cases: 0 ≤ �(0) and 0 ∈≤ �(0). If 0 ≤ �(0) then
x(t) ≥ 0 is the desired solution of the differential inclusion. Let us show that 0 ∈≤ �(0)
implies existence � of a solution q ≤ (0, →), |u| = 1 of the equation f (u) = qu. Indeed, if
0 ∈≤ �(0) and r √ 2 then 0 = ∈ f (� u) for all � ≤ [0, 1], |u| = 1, and hence

f (� u)
(�, u) ∞≈
|f (� u)|
 3

is a homotopy between the vector fields f1 : u ∞≈ f (u)/|f (u)| and f0 : u ∞≈ f (0)/|f (0)|.
Since the index of the constant map f0 is zero, the index of f1 is zero as well. However,
assuming that f (u) ∈= qu for q ≤ (0, →), |u| = 1 yields a homotopy

� u + (1 − � )f (u)
(�, u) ∞≈
|� u + (1 − � )f (u)|

between f1 and the identity map, which is impossible, since the identity map has index 1.
Hence f (u) = qu for some q > 0, |u| = 1, which yields x0 (t) = qtu as as a valid
solution x0 : [0, →) ∞≈ R2 of the differential inclusion.

Problem T1.3
Find all values q, r ≤ R for which x̄0 = 0 is not a (locally) stable equilibrium
of the ODE
ẋ(t) = Ax(t) + B(Cx(t))1/3 (1.1)
for every set of matrices A, B, C of dimensions n-by-n, n-by-1, and 1-by-n
respectively, such that A is a Hurwitz matrix and

Re[(1 + j�q)G(j�)] > r � � ≤ R (1.2)

for
G(s) = C(sI − A)−1 B.

Answer: r √ 0, q ≤ R arbitrary (note, however, that for r √ 0 (1.2) implies q √ 0).

Proof If r < 0, take A = −1, B = 0, C = 1 to get an example of A, B, C satisfying the
conditions and such that x̄0 = 0 is a (globally asymptotically) stable equilibrium of (1.1).
Now consider the case r √ 0. Then, informally speaking, the frequency domain
condition means some sort of “passivity” of G, while (1.1) describes a positive feedback
interconnection of G with nonlinearity y ∞≈ w = y 1/3 , which can be characterized as
having arbitrarily large positive gain for x � 0. Hence one expects instability of the zero
equilibrium of (1.1).
To show local instability, let us prove existence of a Lyapunov function V = V (x) for
which 0 is not a local minimum, and
d
V (x(t)) < 0 whenever |Cx(t)| ≤ (0, τ0 )
dt
for some τ0 > 0. Note that this will imply instability of the equilibrium x̄0 = 0, since
every solution with V (x(0)) < V (0) and |x(0)| < τ0 /2|C | will eventually cross the sphere
|x(0)| = τ0 /2|C | (otherwise |Cx(t)| ∀ τ0 /2 for all t √ 0, hence V (x(t)) is monotonically
 4

non-increasing, and all limit points x¯ of x(·) satisfy C x¯ = 0, therefore every solution x � (t)
of (1.1) beginning at such limit point satisfies Cx� (t) = 0 and hence converges to the
origin, which contradicts V (x(0)) < V (0)).
By introducing w(t) = (Cx(t))1/3 , system equations can be re-written in the form

ẋ(t) = Ax(t) + Bw(t).

Consider first the (simpler) case when r > 0 (and hence q > 0). Then one can use the
inequality
w(t)Cx(t) ∀ r|w(t)|2 ,
for sufficiently small |Cx(t)|. Condition (1.2) together with the KYP Lemma yields exis-
tence of a matrix P = P � such that

wC
¯ x¯ + q wC(A¯
¯ ¯ − r|w¯|2 √ 2¯
x + B w) x� P (A¯ ¯ � x¯ ≤ Rn , w¯ ≤ R.
x + B w)

Substituting w = (Cx)1/3 , we get

d �
[x P x − 0.75q |Cx|4/3 ] ∀ |Cx|4/3 − r|Cx|2/3 ,
dt
�
which is exactly what is needed, because y 4/3 − ry 2/3 < 0 for y ≤ (0, r). In addition, for
every x¯0 ≤ Rn such that C x¯0 ∈= 0 the expression

x) = x¯� P x¯ − 0.75q |C x¯|4/3

V (¯

is negative when x¯ = r¯
x0 and r > 0 is small enough.
To prove the answer in the general case, note that the inequality

w(t)Cx(t) > R|Cx(t)|2

is satisfied whenever |Cx(t)| ≤ (0, τ) with τ > 0 and R = τ−2/3 , i.e. R can be made
arbitrarily large by selecting an appropriate τ > 0. Therefore the derivative bound for
V (x(t)) = x(t)� P x(t) will hold if

x� P (A¯
2¯ x + B w) ¯ x¯ � x¯ ≤ Rn , w¯ ≤ R.
¯ ∀ R|C x¯|2 − wC (1.3)

According to the KYP Lemma, such P = P � exists if

R|G(j�)|2 − Re(G(j�)) > 0 � � ≤ R,

or, equivalently,
Re(1/G(j�)) < R � � ≤ R.
Moreover, substituting w = (R + K)Cx, where K > 0 is a constant, into (1.3) yields

P (A + BKC) + (A + BKC)� P ∀ −KC � C.

 5

Therefore, P = P � cannot be positive semidefinite if A + BKC has eigenvalues with

positive real part.
We will rely on the following statement from the linear system theory: if H(s) is a
stable proper rational transfer function which is positive real (i.e. Re(H(j�)) > 0 for all
� ≤ R) then Re(s) > 0 whenever Re(s) > 0, and the relative degree of H is not larger
than one.
Consider H(s) = (1 + qs)G(s). By assumption, H is positive real and proper. Hence
q √ 0 (otherwise H(−1/q) = 0). If relative degree of H is zero then q > 0, and hence
sG(s) converges to a non-zero limit H� as s ≈ →. Since (1 + qr)G(r) > 0 for r > 0, it
follows that H� > 0, and hence
1 j�
Re = Re
G(j�) G(j�)

is bounded as � ≈ →.
If relative degree of H is one then sH(s) converges to a positive limit as s ≈ →, and
hence
1 j� − (j�)2
Re = Re
G(j�) (j�)2 G(j�)
is bounded from above as � ≈ →. Finally, since G(r) > 0 and G(r) ≈ 0 as r ≈ +→,
it follows that the equation 1 = KG(r) has a positive solution r for all sufficiently large
K > 0. Hence matrix A + BKC has a positive eigenvalue for all sufficiently large K > 0.
 Massachusetts Institute of Technology

Department of Electrical Engineering and Computer Science

6.243j (Fall 2003): DYNAMICS OF NONLINEAR SYSTEMS
by A. Megretski

Take-Home Test 2 Solutions1

Problem T2.1
System of ODE equations

ẋ(t) = Ax(t) + Bψ(Cx(t) + cos(t)), (1.1)

where A, B, C are constant matrices such that CB = 0, and ψ : R k ∈� Rq is

continuously differentiable, is known to have a locally asymptotically
stable non-equilibrium periodic solution x = x(t). What can be said about
trace(A) ? In other words, find the set Λ of all real numbers � such that
� = trace(A) for some A, B, C, ψ such that (1.1) has a locally asymptotically
stable non-equilibrium periodic solution x = x(t).
Answer: trace(A) < 0.
Let x0 (t) be the periodic solution. Linearization of (1.1) around x0 (·) yields

α̇(t) = Aα(t) + Bh(t)Cα(t),

where h(t) is the Jacobian of ψ at x0 (t), and

x(t) = x0 (t) + α(t) + o(|α(t)|).

Partial information about local stability of x0 (·) is given by the evolution matrix M (T ),
where T > 0 is the period of x0 (·): if the periodic solution is asymptotically stable then
all eigenvalues of M (T ) have absolute value not larger than one. Here

Ṁ (t) = (A + Bh(t)C)M (t), M (0) = I,

1
Version of November 25, 2003
 2

and hence ⎛� T ⎝
det M (T ) = exp trace(A + Bh(t)C)dt .
0

Since
trace(A + Bh(t)C) = trace(A + CBh(t)) = trace(A),
det(M (T )) > 1 whenever trace(A) > 0. Hence trace(A) ≈ 0 is a necessary condition for
local asymptotic stability of x0 (·).
Since system (1.1) with k = q = 1, ψ(y) ≥ y,
⎡ � ⎡ �
−a 0 1 � �
A= , B= , C= 0 1
0 −a 0

has periodic stable steady state solution

(1 + a2 )−1 cos(t) + a(1 + a2 )−1 sin(t))

⎡ �
x0 (t) =
0

for all a > 0, the trace of A can take every negative value. Thus, to complete the solution,
one has to figure out whether trace of A can take the zero value.
It appears that the volume contraction techniques are better suited for solving the
question completely. Indeed, consider the autonomous ODE
⎞
⎨ ż1 (t) =
⎨ z2 (t),
ż2 (t) = ⎛ z1 (t),
−
⎠
⎝ (1.2)
z1 (t)
⎧ ż3 (t) = Az3 (t) + Bψ Cz3 (t) +
⎨
⎨ � 2 2
,
z1 (t) +z2 (t)

defined for z12 + z22 ∞= 0. If (1.1) has an asymptotically stable periodic solution x0 = x0 (t)
then, for δ > 0 small enough, solutions of (1.2) with
�⎪ ⎣ � ⎪ ⎣
� z1 (0) � 1
� �
�� z2 (0) ⎤ − z̄ � ≈ δ, z̄ � 0 ⎤
� �
� z3 (0) � x0 (0)

small enough satisfy

lim z3 (t) − x0 (t + β ) = 0,
t→∞

where β � 0 is defined by z2 (−β ) = 0. In particular, the Euclidean volume of the image of

the the ball of radius δ centered at z̄ under the differential flow defined by (1.2) converges
to zero as t � →. Since the volume is non-increasing when trace(A) √ 0, we conclude
that trace(A) < 0.
 3

Problem T2.2
Function g1 : R3 ∈� R3 is defined by
⎩⎪ ⎣� ⎪ ⎣
x1 1
g1 ⎦� x2 ⎤� = � x1 ⎤ .
x3 0

(a) Find a continuously differentiable function g2 : R3 ∈� R3 such that

the driftless system
ẋ(t) = g1 (x(t))u1 (t) + g2 (x(t))u2 (t) (1.3)
is completely controllable on R3 .
For ⎪ ⎣
1
g2 (x) = � 0 ⎤ = const,
1
we have ⎪ ⎣
0
g3 = [g1 , g2 ] = � 1 ⎤ .
0
Since g1 (x), g2 , g3 form a basis in R3 for all x, the resulting system (1.3) is completely
controllable on R3 .
(b) Find continuously differentiable functions g2 : R3 ∈� R3 and h : R3 ∈�
R such that ≡h(¯ x) ∞= 0 for all x¯ ≤ R3 and h(x(t)) is constant on all
solutions of (1.3). (Note: function g2 in (b) does not have to be (and
cannot be) the same as g2 in (a).)
For example, ⎪ ⎣ ⎩⎪ ⎣�
1 x1
g2 (x) = � 1 ⎤ = const, h ⎦� x2 ⎤� = x3 .
0 x3

(c) Find a continuously differentiable function g2 : R3 ∈� R3 such that

the driftless system (1.3) is not completely controllable on R 3 , but,
on the other hand, there exists no continuously differentiable func-
tion h : R3 ∈� R such that ≡h(¯ ∞ 0 for all x¯ ≤ R3 and h(x(t)) is
x) =
constant on all solutions of (1.3).
For ⎪ ⎣
0
g2 (x) = � x1 ⎤ ,
x3
 4

we have ⎪ ⎣
0
g3 = [g2 , g1 ] = � 1 ⎤ ,
0
and hence g1 (x), g2 , g3 form a basis in R3 whenever x3 ∞= 0. This contradicts the
condition that ≡h(x) must be non-zero ad orthogonal to g1 (x), g2 (and hence to g3 )
for all x.

Problem T2.3
An ODE control system model is given by equations
⎞
⎠ ẋ1 (t) = x2 (t)2 + u(t),
ẋ2 (t) = x3 (t)2 + u(t), (1.4)
ẋ3 (t) = p(x1 (t)) + u(t).
⎧

(a) Find all polynomials p : R ∈� R such that system (1.4) is full state
feedback linearizable in a neigborhood of x¯ = 0.
System (1.4) has the form

ẋ(t) = f (x(t)) + g(x(t))u(t), (1.5)

where
x22
⎩⎪ ⎣� ⎪ ⎣ ⎩⎪ ⎣� ⎪ ⎣
x1 x1 1
2
f ⎦� x2 ⎤� = � x3 ⎤ , g ⎦� x2 ⎤� = 1 ⎤.
�
x3 p(x1 ) x3 1

Define
g1 = g, g2 = [f, g1 ], g3 = [f, g2 ], g21 = [g2 , g1 ],
i.e.
4x2 x3 − 2x23
⎩⎪ ⎣� ⎪ ⎣ ⎩⎪ ⎣� ⎪ ⎣ ⎩⎪ ⎣� ⎪ ⎣
x1 2x2 x1 x1 2
g2 ⎦� x2 ⎤� = � 2x3 ⎤ , g3 ⎦� x2 ⎤� = � 2x3 ṗ(x1 ) − 2p(x1 ) ⎤ , g21 ⎦� x2 ⎤� = � 2 ⎤ .
x3 ṗ(x1 ) x3 2x2 ṗ(x1 ) − p̈(x1 )x22 x3 p̈(x1 )

For local full state feedback linearizability at x = 0 it is necessary and sufficient

for vectors g1 (0), g2 (0), g3 (0) to be linearly independent (which is equivalent to
p(0)ṗ(0) = 0) and for g21 (x) to be a linear combination of g1 (x) and g2 (x) for
all x in a neigborhood of x = 0 (which is equivalent to p̈(x1 ) ≥ 2). Hence

p(x1 ) = x21 + p1 x1 + p0 , p0 p1 ∞= 0

is necessary and sufficient for local full state feedback linearizability at x = 0.

 5

(b) For each polynomial p found in (a), design a feedback law

u(t) = K(x1 (t), x2 (t), x3 (t)) = Kp (x1 (t), x2 (t), x3 (t))

which makes the origin a locally asymptotically stable equilibrium

of (1.4).
Since p(0) ∞= 0, x = 0 cannot be made into a locally asymptotically stable equilib-
rium of (1.4). However, the origin z = 0 (i.e. with respect to the new coordinates
z = �(x)) of the feedback linearized system can be made locally asymptotically
stable, as long as 0 ≤ �(Ω) where Ω is the domain of �. Actually, this does not
require any knowledge of the coordinate transform �, and can be done under an
assumption substantially weaker than full state feedback linearizability!
Let
ż(t) = Az(t) + Bv(t) (1.6)
be the feedback linearized equations (1.5), where

z(t) = �(x(t)), x(t) ≤ Ω, v(t) = �(x(t))(u − λ(x(t))).

In other words, let

˙
f (x) = [�(x)] −1 ˙
[A�(x) − B�(x)λ(x)], g(x) = [�(x)] −1
B�(x).

If x¯ ≤ Ω satisfies �(¯
x) = 0 then x¯ is a conditional equilibrium of (1.5), in the sense
that
f (¯
x) + g(¯ u=0
x)¯
for u¯ = λ(¯ x). Moreover, since the pair (A, B) is assumed to be controllable, the
conditional equilibrium has a controllable linearization, in the sense that the pair
(f˙(¯
x) + ġ(¯
x)¯ x)) is controllable as well, because
u, g(¯

f˙(¯
x) + ġ(¯ u = S −1 (AS − BF ), g(¯
x)¯ x) = S −1 B�(¯
x)

for
S = �˙ (¯
x), F = �(¯
x)λ̇(¯
x).
It is easy to see that every conditional equilibrium x¯ of (1.5) with a controllable lin-
earization can be made into a locally exponentially stable equilibrium by introducing
feedback control
u(t) = u¯ + K(x(t) − x),
¯
where K is a constant gain matrix such that

f˙(¯
x) + ġ(¯ u + g(¯
x)¯ x)K
 6

is a Hurwitz matrix. Indeed, by assumption x̄ is an equilibriun of

ẋ(t) = fK (x) = f (x(t)) + g(x(t))(¯

u + K(x(t) − x)),
¯

and
f˙K (¯
x) = f˙(¯
x) + ġ(¯ u + g(¯
x)¯ x)K.

In the case of system (1.4) let ⎪ ⎣

x̄1
x¯ = � x¯2 ⎤
x̄3
be a conditional equilibrium, i.e.

x¯21 = x¯22 = p(¯

x1 ) = −¯
u.

Then ⎪ ⎣ ⎪ ⎣
0 x¯2 0 1
f˙(¯
x) + ġ(¯ u=� 0
x)¯ 0 2¯ x) = � 1 ⎤ .
x2 ⎤ , g(¯
ṗ(x̄1 ) 0 0 1
Hence a locally stabilizing controller is given by

x21 + k1 (x1 (t) − x¯1 ) + k2 (x2 (t) − x¯2 ) + k3 (x3 (t) − x¯3 ),
u(t) = −¯

where the coefficients k1 , k2 , k3 are chosen in such a way that

⎪ ⎣ ⎪ ⎣
0 x̄2 0 1 � �
� 0 0 2x̄2 + 1 ⎤ k1 k2 k3
⎤ �
ṗ(x̄1 ) 0 0 1

is a Hurwitz matrix.

(c) Find a C ∞ function p : R ∈� R for which system (1.4) is globally full

state feedback linearizable, or prove that such p(·) does not exist.
Such p(·) does not exist. Indeed, otherwise vectors
⎪ ⎣ ⎪ ⎣
1 2x2
� 1 ⎤ and � 2x3 ⎤
1 ṗ(x1 )

are linearly independent for all real x¯1 , x¯2 , x¯3 , which is impossible for

x¯2 = x¯3 = 0.5p(¯

˙ x1 ).