Bounded-Real Lemma based H∞ synthesis and LQR

synthesis for discrete-time polytopic systems

November 7, 2021

1 Polytopic system:
Consider the following discrete-time polytopic system:
N
!
X
xk+1 = αk (i)Ai xk + Buk + Ewk (1)
i=1

where xk ∈ Rn is the state vector, at instant k, and xk+1 its successor, uk ∈ Rp and wk ∈ Rm
are the control input and the process disturbance input, respectively. Matrices Ai with
i = {1, 2, . . . , N }, B and E are the state-space matrices with appropriated dimensions.

LPV systems where the scheduling parameter ρk appears in an affine form into the state-
space matrices, can be written in a polytopical form as follows:
N
!
X
A(ρk ) = αk (i)Ai (2)
i=1

for a given scheduling parameter ρk ∈ Ωρ , then there exists a column vector αk ∈ RN , formed
by positive scalar elements αk (i) ≥ 0, i = {1, . . . , N }, such that
 
αk (1)
θ1 θ2 . . . θN  αk (2) 
    
ρk
.. = (3)
1 1 ... 1  1

. 
| {z }
M (θ) αk (N )
| {z }
αk

where the column vectors θi ∈ RL , i = {1, . . . , N }, are the a priori known vertices of the
polytopic set Ωρ ⊂ RL . Notice that the column vector αk ∈ RN , solving the Problem (3) can
be obtained as a solution of a LP problem. In Section ?? it is presented an algorithm for
solving this problem in real-time, allowing the implementation of polytopic controllers and/or
observers.

1
2 Polytopic H∞ state feedback control design
For a given symmetric matrix Q ∈ Rn×n , positive definite Q  0. The closed-loop system has a
quadratic H∞ performance less than γ (γ > 0) if there exist a symmetric positive definite matrix
U ∈ Rn×n and matrices Yi ∈ Rn×p , for i = {1, . . . , N }, satisfying the following condition:

UAi − Yi BT
 
−U 0n×m U
 ? −γ 2 Im×m ET 0m×n  0
 (4)
 ? ? −U 0n×n 
? ? ? −Q−1

for i = {1, . . . , N }.

and the following control input, uk ∈ Rp , is performed as

N
!
X
uk = − αk (i)Ki xk (5)
i=1

where the polytopic state-feedback gain matrices , Ki ∈ Rp×n , are obtained as:

Ki = (U−1 Yi )T (6)

for i = {1, . . . , N }.

In terms of performance, we can say that, for energy-bounded disturbances:

xTk Qxk ≤ γ 2 wTk wk (7)

This means that we can chose a matrix Q to focus the optimization to a particular ”perfor-
mance” output. For instance, if we are interested on minimizing the impact of the disturbance
wk on the output zk , with zk = Hxk , then we can chose Q = HT H and obtaining that:

zTk zk ≤ γ 2 wTk wk (8)

3 Polytopic H∞ state observer design

Consider the previous polytopic system (1), with measurements yk ∈ Rq described by:

yk = Cxk + Fvk (9)

where vk stands by measurements noise. Consider the polytopic estimation error dynamics:
N
!  
X wk
ek+1 = αk (i)(Ai − Li C) ek + [E Li F] (10)
vk
i=1

For a given symmetric matrix Q ∈ Rn×n , Q  0, the H∞ norm of the estimation error
dynamics is less than γ > 0 if there exist a symmetric positive definite matrix P ∈ Rn×n and
matrices Ui ∈ Rn×q satisfying the following condition:
ATi P − CT UTi
 
−P + Q 0n×m
 ? −γ 2 Im×m [PE − Ui F]T   0 (11)
? ? −P

2
for i = {1, . . . , N }.
Moreover, the observer gain matrices Li ∈ Rn×q , can be obtained as

Li = P−1 Ui (12)

for i = {1, . . . , N }.
The polytopic observer is implemented as follows:
N
! N
!
X X
x̂k+1 = αk (i)Ai x̂k + Buk + αk (i)Li (yk − Cx̂k ) (13)
i=1 i=1

In terms of performance, we can say that, for energy-bounded disturbances:

eTk Qek ≤ γ 2 dTk dk (14)

 
wk
with dk = .
vk
This means that we can chose a matrix Q in (11) to focus the optimization to a particular
estimation error output. For instance, if we are interested on minimizing the impact of the
T
disturbance dk on the output z̃k , with z̃k = H̃ek , then we can chose Q = H̃ H̃ and obtaining
that:
z̃Tk z̃k ≤ γ 2 dTk dk (15)
We can also use projection of ellipsoids to find the H∞ gain performance for a given system
output described as a linear coordinate transformation of the full state (see next subsection).

3.1 Projection of ellipsoids

Consider any linear coordinate transformation, for instance zk = Cxk , then, for any state xk
belonging to the set:

εx = {xk ∈ Rn | xTk Σ−1 xk ≤ 1} (16)

the output zk will be inside the set:

εz = {zk ∈ Rm | zTk (CΣCT )−1 zk ≤ 1} (17)

with m ≤ n.

3
4 Polytopic LQR synthesis
A LQR problem can be solved by finding a state feedback gain K and a matrix P  0 such
that:
min trace (C − DK)P (C − DK)T


(18)
P >0 (19)
(A − BK)P (A − BK)T − P + Q ≤ 0 (20)
which means that we looking for the gain K which minimizes a quadratic function of the output
z = Cx + Du (i.e. the norm 2 of z) with
 1/2 
Q
C= (21)
0
and
 
0
D= (22)
R1/2
The problem (18)-(20) can be numerically solved by re-written the problem as a LMI. Remark
that condition (20) can be written as:
(A − BK)P P −1 P (A − BK)T − P + Q ≤ 0 (23)
Now applying a Schur complement, the condition becomes:
 
−P + Q (AP − BKP )
0 (24)
∗ −P
Thus, by considering a change of variables, Y := KP :
 
−P + Q (AP − BY )
0 (25)
∗ −P
In addition, the cost function has to be rewritten by defining an upper bound as follows:

W ≥ (C − DK)P (C − DK)T


(26)
and by applying a Schur complement to that condition, we have:
 
W (CP + DY )
0 (27)
∗ P
The LQR problem becomes to find W , P and Y such that
min trace (W ) (28)
P >0 (29)
subject to condition (25) and (27).

The state feedback gain will computed as K = Y P −1 .

Exercise: Explain why it is possible to use that problem statement to solve the LQR
problem for polytopic systems ?
Write the problem in a polytopic form. Provide an example of Matlab code which allow us to
numerically solve this problem (use for instance CVX toolbox).