Professional Documents
Culture Documents
To cite this article: Adam C. Sniderman, Mireille E. Broucke & Gabriele M. T. D'Eleuterio (2018)
Control of a class of patterned systems, International Journal of Control, 91:7, 1489-1507, DOI:
10.1080/00207179.2017.1319578
Article views: 64
The framework of patterned linear systems was outcome was obtained for patterned linear systems with
extended to multivariate agents by Sniderman, Broucke, scalar agents.
and D’Eleuterio (2013, 2015a), Consolini and Tosques While this paper is directly related to those of Hamil-
(2014), and Holmes and Broucke (2016). To realise this ton and Broucke (2012), Sniderman et al. (2013, 2015a),
extension, a new algebraic encoding of the system pat- Consolini and Tosques (2014), and Holmes and Broucke
tern was adopted in terms of commuting relationships of (2016), our work also finds roots in earlier research study-
the system matrices, rather than polynomials of a base ing systems whose matrices satisfy commuting relation-
matrix. First, Sniderman et al. (2013) studied systems ships. First, Hazewinkel and Martin (1983) contributed
with the particular block circulant pattern, and showed early work involving commuting relationships. Inspired
that many standard control problems could be solved by by a seminal paper on block circulant systems (Brock-
block circulant feedbacks under the same necessary and ett & Willems, 1974), they studied a more general class
sufficient conditions. To extend the class of block circu- of systems with symmetries, and characterised those sys-
lant systems, Sniderman et al. (2015a) looked at patterns tems’ controllability and stabilisability in terms of sub-
encoded by any single commuting relationship between systems determined via the symmetry algebra, producing
the system matrices and a base matrix, and showed that results similar in style to those of Holmes and Broucke
the standard notion of controllability does not suffice (2016). Second, Fagnani and Willems (1994) developed a
to allow for pole placement with a patterned feedback method to embed symmetries into the behavioural con-
matrix, in contrast with classical pole placement in linear trol approach, and found that symmetric controllers exist
systems. Rather, patterned pole placement operates under whenever standard controllers exist. Our work is comple-
the necessary and sufficient condition of the controllabil- mentary to the above two, and differs in two key ways: our
ity of a so-called ‘reduced system’. Consolini and Tosques patterns involve a diagonalisable base matrix, while the
(2014) encoded patterns using the automorphism group above symmetries use several diagonalisable and invert-
of a certain system graph for systems whose patterns ible matrices; and we study several geometric control
are completely described by symmetries. These symme- problems other than pole placement and stabilisation,
tries yield a family of commuting relationships between which are not explicitly covered in prior works.
the system matrices and a (matrix) representation of the This paper is organised as follows. Section 2 con-
automorphism group of the graph. However, a patterned tains standard background on linear algebra and linear
pole placement theorem was not postulated, and there- systems, as well as notation. Section 3 introduces pat-
fore only a subset of synthesis problems could be solved. terned matrices, and lays out the main mathematical
Holmes and Broucke (2016) adopted the framework of tools for keeping track of patterns when solving control
Consolini and Tosques (2014) to solve the Patterned Pole problems. First, we discuss the algebra of patterns in
Placement Problem. The approach followed that of Sni- terms of commuting relationships with certain matrices,
derman et al. (2015a) to study the controllability of a as well as the geometry of patterns in terms of certain
reduced system (which effectively quotients out the graph invariant subspaces under those matrices. To tie the
symmetries). algebra and geometry together, we show how these
In this paper, we follow the framework of Sniderman commuting relationships carry through restrictions to
et al. (2015a) to consider a class of patterned systems these invariant subspaces, thereby providing a way to
where the system matrices satisfy a single commuting preserve patterns in the state space decompositions
relationship with a base matrix that is diagonalisable. of geometric control. Then, we discuss the spectra of
A key question arises: do the standard results of linear patterned matrices, and examine patterned solutions
control theory apply to this class, without requiring addi- to linear equations of patterned matrices. With this
tional conditions (as found by Sniderman et al. (2015a) machinery in hand, we turn to patterned control systems
and Holmes and Broucke (2016))? We answer this ques- in Section 4. After defining initial concepts, Sections 5
tion affirmatively for patterned systems whose base and 6 discuss what the fundamental properties of con-
matrices are diagonalisable. With this class of patterned trollability and observability mean for patterned systems.
systems, the difficulties encountered by Sniderman et al. Section 5 also presents methods for pole placement
(2015a ) and Holmes and Broucke (2016) to achieve a and stabilisation of patterned systems using patterned
patterned pole placement theorem do not arise. Instead, feedbacks. Using these techniques, Sections 7, 8, 9
we arrive at the appealing outcome that all the standard and 10 study several other multivariable controller
results of linear control theory carry over to this class of synthesis problems on patterned systems, including
patterned linear systems without any changes in the nec- the Output Stabilisation Problem (OSP), Disturbance
essary and sufficient conditions. This finding echoes the Decoupling Problem (DDP), Stabilisation by Mea-
results of Hamilton and Broucke (2012), where a similar surement Feedback Problem (SMFP) and Restricted
INTERNATIONAL JOURNAL OF CONTROL 1491
Regulator Problem (RRP). Section 11 provides a the top-right block ‘*’ will be 0 – and Equations (1) and
numerical example to demonstrate the patterned synthe- (2) will also hold for S2 and Q1 .
sis technique for a pattern that occurs often in distributed Let C = Cm Cc be a disjoint partition of the com-
systems. Section 12 gives concluding remarks to further plex plane (so Cm ∩ Cc = ∅). Split the minimal polyno-
assess the significance of the findings. mial of A as ψ(s) = ψ m (s)ψ c (s) according to its fac-
tors in Cm and Cc . The corresponding modal subspaces
are X m (A) := Ker ψ m (A) and X c (A) := Ker ψ c (A). The
2. Mathematical preliminaries following result relates this modal decomposition to other
This section presents some mathematical background state space decompositions of X .
and notation. Given a matrix A, denote its transpose by Lemma 2.1 (Wonham, 1979): Let A : X → X be a linear
AT and its complex conjugate by Ā. If A is square, denote map. Let S ⊂ X be an A-invariant subspace, and define A1
its spectrum by σ (A); also denote its ‘distinct spectrum’ and A2 as above. Let C = Cm Cc be a disjoint partition
– the set containing one of each distinct eigenvalue – by of C. Then, X m (A) ⊂ S if and only if σ (A2 ) ⊂ Cc , and
σd (A) ⊂ σ (A). Given a vector space X with indepen- X m (A) ∩ S = 0 if and only if σ (A1 ) ⊂ Cc .
dent subspaces R, S ⊂ X (so R ∩ S = {0}), denote their
direct sum by R ⊕ S.
It is assumed that the reader is already familiar with the 3. Patterned matrices and decoupling
tools of linear geometric control theory (Basile & Marro, subspaces
1991; Wonham, 1979). Let X be a vector space with Our framework for control of patterned systems is based
subspace S ⊂ X and complementary subspace S c ⊂ X , on an interplay between the algebraic properties of com-
so X = S ⊕ S c . Geometric control involves state space muting matrices and the geometric properties of invari-
decompositions that are based on two standard projec- ant subspaces. In this section, we introduce some of the
tion maps: the insertion and the natural projection. The main tools for solving the problem of patterned control
insertion map on S, denoted S : S → X , maps x ∈ S to synthesis.
the corresponding element x ∈ X ; that is, Sx x. The Let F ∈ {C, R} be a field, and define matrices V ∈
natural projection on S along S c , denoted Q : X → S, Fn×n and U ∈ Fm×m .
maps x ∈ X to its component in S; that is, using the
Definition 3.1: A matrix A ∈ Fn×m is called a patterned
unique decomposition x = s + r with s ∈ S and r ∈ S c ,
matrix if it satisfies the commuting relationship VA = AU.
Qx s. These maps satisfy QS = IS , where IS is the iden-
This relationship is denoted by A ∈ C(V, U ), and V and
tity map on S. Also, S = Im S and S c = Ker Q.
U are called the base matrices of the pattern. If V = U,
Let A : X → X be a linear map, and suppose that
then we use the shorthand C(V ) := C(V, V ). If the field
S is A-invariant; that is, AS ⊂ S. The restriction of A
is important, we write CF (V, U ).
to S, denoted by the linear map AS : S → S, is the
unique solution of the equation AS = SAS ; thus, AS per- A pattern prescribes a specific form in a matrix, and so
forms the action of A on S, and is not defined off S. is useful for encoding a number of common distributed
In a basis for
X adapted to S, A has the matrix rep- system structures. We present two particular examples
A1 ∗
resentation 0 A2 , where A1 is a matrix representation here, a ring and a bidirectional chain, which fit into the
of AS (denoted A1 ≡ AS ). This matrix representation model used in this paper.
can be produced by the above projection maps: define Example 3.1 (Ring): Consider a system made up of three
insertions and natural projections on S by S1 , Q1 and parts, whose interconnections form a ring (as in Figure
on S c by S2 , Q2 , and define the coordinate transforma- 1 (left)). The parts are identical, each having the same n
tion T = S1 S2 . Then, T −1 = col(Q1 , Q2 ), and T−1 AT states and m inputs, and the interconnections among the
yields a matrix with the above form, where A1 = Q1 AS1
and A2 = Q2 AS2 . The following relationships can be
deduced: 2 1
AS1 = S1 A1 (1) 1 2
Q2 A = A2 Q2 . (2) 3 3
If S c is also A-invariant, then the matrix representa- Figure . Distributed systems with diagonalisable base matrices:
tion in the adapted basis will be block diagonal – i.e. ring (left) and bidirectional chain (right).
1492 A. C. SNIDERMAN ET AL.
parts are also identical; for example, the coupling between in that they fully decouple a state space into two invariant
parts ‘1’ and ‘2’ is the same as the coupling between parts parts. Invoking these decoupling subspaces is the crucial
‘2’ and ‘3’, and between ‘3’ and ‘1’. This manifests as a block step that ties together the algebraic properties of commut-
structure in all the system matrices, made up of diagonal ing matrices and the geometric notions of linear geomet-
bands of identical blocks (wrapped around cyclically): ric control (Basile & Marro, 1991; Wonham, 1979).
⎡ ⎤ Definition 3.2: A subspace S ⊂ X is called V-
M1 M2 M3 decoupling if it is V-invariant, and it has a V-invariant
M = ⎣ M3 M1 M2 ⎦, Mi ∈ Fn×m . complement S c . That is, V S ⊂ S and V S c ⊂ S c .
M2 M3 M1
In general, a V-invariant subspace is not V-decoupling.
This block structure is called block circulant, and can be However, if V is diagonalisable, then this will be the case
uniquely described by a commuting relationship involv- (Gohberg, Lancaster, & Rodman, 1986, Theorem 3.2.1).
ing the fundamental permutation matrix For this reason, we focus only on patterns with diag-
⎡ ⎤ onalisable base matrices, and always assume that this
0 1 0 property holds. Such an assumption is not unfounded; it
3 = ⎣ 0 0 1 ⎦. appears in a number of other recent papers (Consolini &
1 0 0 Tosques, 2014, 2015; Deutscher, 2013; Massioni & Ver-
haegen, 2009) (and is less restrictive than some of them).
Specifically, M is block circulant if and only if In fact, some such restriction appears in all work to date
(3 In )M = M(3 Im ) (where ‘’ represents the regarding control with commuting relationships, except
Kronecker product), and it follows that every ring system for some preliminary results in Sniderman et al. (2015a).
is (3 In )-patterned. Assumption 3.1: Throughout the paper (even when not
Example 3.2 (Bidirectional chain): Consider a system explicitly stated), we assume that all base matrices are diag-
made up of three parts, whose interconnections form a onalisable. For instance, if M ∈ C(V, U ), then V and U are
bidirectional chain (as in Figure 1 (right)). Again, the diagonalisable.
parts are identical and have identical interconnections.
This again manifests as a block structure in all the system The next two results state that decoupling and diago-
matrices; one possible such structure is given by nalisability are maintained in common subspace opera-
tions and restrictions.
⎡ ⎤
M1 M2 M3 Lemma 3.1 (Gohberg et al., 1986): Let V be diagonalis-
M = ⎣ M2 M1 + M3 M2 ⎦, Mi ∈ Fn×m , able and let S ⊂ X be V-invariant (or -decoupling). Then
M3 M2 M1 VS is diagonalisable.
Lemma 3.2: Let V be a diagonalisable matrix, and let
where the coupling in the middle term is in the style S1 , S2 ⊂ X be V-decoupling subspaces. Then, S1 + S2
of a mass-spring system. This block structure can be and S1 ∩ S2 are also V-decoupling subspaces.
uniquely described by a commuting relationship involv-
Proof: Since S1 and S2 are V-invariant, therefore S1 + S2
ing the matrix
and S1 ∩ S2 are also V-invariant (Gohberg et al., 1986,
⎡ ⎤ Proposition 1.8.1). Since V is diagonalisable, therefore
0 1 0
⎣ S1 + S2 and S1 ∩ S2 also have V-invariant complements
Ṽ = 1 0 1 ⎦.
(Gohberg et al., 1986, Theorem 3.2.1), so they are V-
0 1 0
decoupling by definition.
Specifically, M has the above block structure if and only Next, consider A ∈ C(V ), and let S be an A-invariant
if (Ṽ ⊗ In )M = M(Ṽ ⊗ Im ) (where ‘’ denotes the Kro- subspace. Is S also V-invariant (or V-decoupling)? Not
necker product), and it follows that every bidirectional generally. Fortunately, several specific A-invariant sub-
chain system is (Ṽ ⊗ In )-patterned. spaces are V-decoupling, including those needed to solve
Knowing that many common distributed systems can the control problems in this paper.
be encoded by patterns, the central question of this Lemma 3.3: Let A ∈ C(V ), B ∈ C(V, U ) and K ∈
paper is how to keep sight of a pattern while perform- C(U, V ). Let ρ be a polynomial, and let S ⊂ X be a V-
ing the standard decompositions and syntheses for con- decoupling subspace. Then,
trol design. The key idea in linking these notions lies in
finding suitable invariant subspaces that also have invari- (i) Im B and Ker K are V-decoupling subspaces.
ant complements. These subspaces are called decoupling, (ii) Im ρ(A) and Ker ρ(A) are V-decoupling subspaces.
INTERNATIONAL JOURNAL OF CONTROL 1493
(iii) AS (the image of S under A) and A−1 S := {x ∈ Proof: Let {x1 , … , xk , xk + 1 , … , xn } be a pre-
X | Ax ∈ S} (the preimage of S under A) are V- ferred basis for X , so S = span{x1 , . . . , xk } and
decoupling subspaces. S c = span{xk+1 , . . . , xn }. From this basis, define
the insertions and natural projections S1 : S → X ,
Proof: S2 : S c → X , Q1 : X → S and Q2 : X → S c . The coor-
dinate transformation T = [ S1 S2 ] gives the standard
(i) Let x ∈ Im B, so x = Bv for some v. Then, V x = representation
V (Bv ) = B(U v ) ∈ Im B. Let x ∈ Ker K, so Kx =
0. Then, K(Vx) = U(Kx) = 0, so V x ∈ Ker K. −1 Q1 AS1 ∗
T AT = .
(ii) ρ(A) ∈ C(V ) by Lemma A.1, so Im ρ(A) and 0 Q2 AS2
Ker ρ(A) are V-invariant by (i).
(iii) First, V (MS ) = M(V S ) ⊂ MS, so MS is V- Then, A1 := Q1 AS1 = Q1 S1 AS = AS (using Equation
invariant. Second, let x ∈ M −1 S. Since Mx ∈ S (1)) and A2 Q2 AS2 , giving Equation (3). Further, A1 ∈
and V S ⊂ S, it follows that M(V x) = V (Mx) ∈ C(VS ) and A2 ∈ C(VS c ) by Lemma 3.4(i).
S, so V x ∈ M −1 S. Thus, M −1 S is V-invariant. Lemma 3.4 and Theorem 3.1 start with patterned
matrices and show that their restrictions have patterns
involving the restrictions of the base matrices. Con-
Having stated the main properties of decoupling sub-
versely, we next start with matrices that commute with
spaces, we next explore their utility in maintaining pat-
these restrictions, and show that they can be lifted back
terns in restrictions of patterned matrices. In particular,
up to full patterned matrices.
if A is a V-patterned matrix, and S is an A-invariant and
V-decoupling subspace, then the restriction AS will com- Lemma 3.5 (Patterned Lifting Lemma): Let S ⊂ X be
mute with the restriction VS . a V-decoupling subspace with insertion S1 : V → X and
natural projection Q1 : X → V (along a V-invariant com-
Lemma 3.4: Let S1 , S2 ⊂ X be V-decoupling subspaces. plement).
Let S1 , Q1 and S2 , Q2 be the insertions and natural projec-
tions on S1 and S2 , respectively. Also define the restrictions (i) If A1 ∈ C(VS ), then S1 A1 Q1 ∈ C(V ).
V1 := VS1 and V2 := VS2 . For i, j = 1, 2, (ii) If B1 ∈ C(VS , U ), then S1 B1 ∈ C(V, U ).
(iii) If K1 ∈ C(U, VS ), then K1 Q1 ∈ C(U, V ).
(i) If A ∈ C(V ), then Qi AS j ∈ C(Vi , V j ).
(ii) If B ∈ C(V, U ), then Qi B ∈ C(Vi , U ). Proof: Using Equations (1) and (2),
(iii) If K ∈ C(U, V ), then KSi ∈ C(U, Vi ).
(S1 A1 Q1 )V = S1 A1VS Q1 = S1VS A1 Q1 = V (S1 A1 Q1 )
Proof: Using Equations (1) and (2), (Qi ASj )Vj =
Qi AVSj = Qi VASj = Vi (Qi ASj ), giving (i). The proofs of giving (i). The proofs of (ii) and (iii) are similar.
(ii) and (iii) are similar.
Transitioning between a full space and a subspace is
The next result shows how matrix patterns are pre- one of the main ideas in linear geometric control. The
served through state space decompositions. This result ability to perform these transitions without sacrificing
exploits the fact that the restriction of a patterned matrix commuting relationships, as shown by Theorem 3.1 and
to a V-decoupling subspace retains the essential algebraic Lemma 3.5, provides the method by which feedback laws
property of commuting with certain matrices, an imme- developed through linear geometric control will be able
diate result of Lemma 3.4(i). Thus, commuting becomes to preserve a control system’s pattern. The matrices that
the key enabling property in patterned decompositions. occur in these decompositions will always involve restric-
tions of V-patterned matrices to V-decoupling subspaces.
Theorem 3.1 (Patterned Representation Theorem): Let On the other hand, we have already mentioned that
A ∈ C(V ), and let S ⊂ X be an A-invariant and V- block diagonalisation is also an important factor in con-
decoupling subspace, with V-invariant complement S c ⊂ trol syntheses, and is still required for two steps in our
X . Then, A has a matrix representation framework. Block diagonalisation of patterned matri-
ces is based on the Jordan forms of its base matrices
A1 ∗ via a classic result in Gantmacher (1959), stated here as
(3)
0 A2 Theorem A.1. Block diagonalisation has been the central
idea of most previous research (e.g. Brockett & Willems,
where A1 ≡ AS . Moreover, A1 ∈ C(VS ) and A2 ∈ C(VS c ). 1974; Hovd & Skogestad, 1994; Hovd et al., 1997; Lunze,
1494 A. C. SNIDERMAN ET AL.
1986; Massioni & Verhaegen, 2009 for block circulant and The most important result on block diagonalisation,
related systems): these papers start by block diagonalising explaining how a real patterned matrix can be turned into
the system, then designing a feedback law on each of the a block diagonal matrix and vice versa, is given in the
blocks, and finally transforming back to original coordi- statement below. The proof can be found in the appendix,
nates. A novelty of this work is that our design approach along with more supporting results of this type.
begins with the standard control theoretic decomposi- Lemma 3.6: Let V and U be real and follow the ordering
tions on the full system, whereas block diagonalisation is
of Remark 3.1. Let matrices A and  satisfy  = V−1 AU .
suppressed to the lowest level step in control design. This
Then, A ∈ CR (V, U ) if and only if  = diag(Â1 , . . . , Âr ),
approach is intuitively appealing as it decouples the con- ¯
where Âi ∈ Cn ×m and Âε = Âi for each i = 1, … , r.
i i i
trol problem from the pattern – it allows standard tech-
niques to be used without consideration of the system’s
An application of the above result is to find patterned
pattern. Block diagonalisation still remains the mecha-
solutions to linear equations of patterned matrices. The
nism by which we explicitly synthesise patterned con-
proof is in the Appendix.
trollers, specifically to solve the Patterned Pole Placement
Problem and patterned linear equations. Next, we present Lemma 3.7: Suppose A ∈ C(V, U ) and B ∈ C(W, U )
a particular ordering of the eigenvalues of V and U that (with W, V and U all diagonalisable, following
facilitates block diagonalisation. Assumption 3.1). Also suppose that the equation XA = B
has a solution X. Then, there exists a patterned solution
Remark 3.1 (Ordering conventions): Throughout the
X ∈ C(W, V ).
paper, superscript indices refer to items ordered accord-
ing to this remark, and subscript indices refer to items At this point, the foundations have been laid in the
with no particular ordering. Let V and U be diagonalis- areas of patterned matrices, decoupling subspaces and
able, and denote the set of their distinct eigenvalues by restrictions to subspaces and lifts to the full state space
σd (V ) ∪ σd (U ) = {δ 1 , . . . , δ r }, where each δ i has multi- of patterned matrices. Patterned control systems can now
plicity ni in σ (V ) and mi in σ (U ). (Note that some ni or be studied from a geometric viewpoint.
mi may be zero, reflecting that δ i does not appear in the
corresponding spectrum.) Order all the eigenvalues of V
and U as follows: 4. Patterned systems
n1 times n times
r Consider the linear time-invariant system given by
1
1
σ (V ) = δ , . . . , δ , . . . , δ , . . . , δ r ,
r
ẋ(t ) = Ax(t ) + Bu(t ) + Ew(t )
σ (U ) = δ 1 , . . . , δ 1 , . . . , δ r , . . . , δ r . y(t ) = Cx(t )
(6)
m1 times mr times
z(t ) = Dx(t )
Next, partition V and U in the same way, so
where x(t ) ∈ Rn is the state vector, u(t ) ∈ Rm is the input
V = V1 ··· Vr , U = U1 ··· Ur (4) vector, w(t ) ∈ Rs represents a disturbance, y(t ) ∈ R p is
the measurement vector and z(t ) ∈ Rq is the output vec-
where each Vi (ni columns) and Ui (mi columns) are tor. The state space, input space, measurement space and
constructed from eigenvectors corresponding to eigen- output space are denoted by X , U, Y and Z, respec-
value δ i ; these full sets of eigenvectors exist because V and tively. Assume a real system with matrices A ∈ Rn×n , B ∈
U are diagonalisable (Gohberg et al., 1986). Finally, diago- Rn×m , C ∈ R p×n , D ∈ Rq×n and E ∈ Rn×s . If all matri-
nalise V and U as V̂ := V−1V V = diag(δ 1 In1 , . . . , δ r Inr ) ces in the system are patterned – that is, A ∈ C(V ), B ∈
and Û := U−1U U = diag(δ 1 Im1 , . . . , δ r Imr ), so the C(V, U ), C ∈ C(Y, V ), D ∈ C(Z, V ) and E ∈ C(V, W ) –
eigenvalues appear in the same order. then Equation (6) is called a V-patterned system.
If V and U are real, then their spectra are symmetric, so The open-loop poles of the system are the eigenval-
there exists a ‘conjugate permutation’ {ε1 , … , εr } of {1, … , ues of A; for patterned systems, they form a so-called
r} such that for all i = 1, … , r, V-patterned spectrum, explained below. The stable and
unstable subspaces of A are denoted X − (A) and X + (A).
δ ε = δ̄ i , nε = ni , mε = mi .
i i i
(5) Finally, we refer to a system as unpatterned if no commut-
ing relationships are explicitly being considered – that is,
Choose the eigenvectors of V and U to follow this con- if the system matrices’ elements are not explicitly con-
jugate permutation, so Vε = ¯ Vi and Uε = ¯ Ui for each
i i
strained by certain algebraic relationships, meaning that
i = 1, … , r. any controller would be a valid choice for the system.
INTERNATIONAL JOURNAL OF CONTROL 1495
The duality between controllability and observability By definition, (A, C) is patterned detectable.
(Wonham, 1979, Lemma 3.4) gives the following result (Only If): The solvability condition is identical to that
about the observability of patterned systems, following for general stabilisability. Since this condition is neces-
immediately from Theorem 5.1. sary for the existence of any feedback, it is also neces-
sary for the existence of a patterned feedback in the class
Theorem 6.1: Let (C, A) be a patterned system satisfying
C(U, V ).
Assumption 3.1. Then, (C, A) is observable if and only if
for every V-patterned spectrum L, there exists a patterned
feedback K ∈ C(V, Y ) such that σ (A + KC) = L. 7. Output stabilisation
Consider the system (D, A, B) from Equation (6), where
6.1 Observable decomposition D corresponds to the system output z(t). The Output Sta-
bilisation Problem (OSP) is to find a state feedback u(t)
Any system can be transformed using a basis that sep-
= Kx(t) such that z(t) → 0 as t → ; in geometric terms,
arates its unobservable and observable parts. If the sys-
X + (A + BK ) ⊂ Ker D. The OSP can be stated equiva-
tem is patterned, then each of these parts will also be
lently for patterned systems with patterned feedback, as
patterned – the resulting commuting relationships are
given in Problem 7.1, and this section shows that the stan-
found using the Patterned Representation Theorem 3.1,
dard OSP and the Patterned OSP are equally solvable for
as given in the following Second Decomposition Theo-
patterned systems; that is, if any output-stabilising feed-
rem. The proof is by duality with the First Decomposition
back exists, then a patterned output-stabilising feedback
Theorem 5.2.
exists.
Theorem 6.2 (Second Decomposition Theorem): Let Problem 7.1 (Patterned OSP): Let (D, A, B) be a pat-
(C, A) be a patterned system satisfying Assumption 3.1, terned system satisfying Assumption 3.1. Find a pat-
and let its unobservable subspace be N with V-invariant terned state feedback K : X → U, K ∈ C(U, V ), such
complement N c . Then, there exists a coordinate trans- that X + (A + BK ) ⊂ Ker D.
formation T : X → X such that the transformed system
matrices have the form Solving the OSP requires the notion of controlled
invariant subspaces: a subspace V ⊂ X is controlled
invariant if there exists F : X → U such that (A +
−1 A1 ∗
T AT = , CT = 0 C2 (9) BF )V ⊂ V; then, F is called a friend of V. The set of
0 A2
all controlled invariant subspaces in X is denoted by
where A1 ≡ AN ∈ C(VN ), A2 ∈ C(VN c ) and C2 ∈ I(X ). The notions of controlled invariance and friends
C(Y, VN c ). Moreover, the pair (C2 , A2 ) is observable. carry over to patterned systems: the next result shows
that it is always possible to find a patterned friend for a
V-decoupling controlled invariant subspace.
6.2 Detectability Lemma 7.1: Let (A, B) be a patterned system. If V ⊂ X is
A patterned system (C, A) is detectable by patterned a V-decoupling subspace, and V ∈ I(X ), then there exists
feedback if there exists a patterned matrix K for which F ∈ C(U, V ) such that (A + BF )V ⊂ V.
σ (A + LC) ⊂ C− . Detectability by patterned feedback Proof: Suppose dim V = k ≤ n. Let W be a V-invariant
can be determined by the same condition as standard complement of V. Let S1 , Q1 and S2 , Q2 be the insertion
detectability. The proof follows by considering the dual and natural projection maps on V and W, respectively.
of Theorem 5.3. Since V is a controlled invariant subspace, AV ⊂ V + B
Theorem 6.3 (Patterned Detectability): Let (C, A) be a (Wonham, 1979 Lemma 4.2). This equation can be writ-
patterned system satisfying Assumption 3.1. Then, (C, A) ten in matrix form as
is detectable by a patterned state feedback K ∈ C(V, Y ) if
AS1 = P − BR, (10)
and only if N ⊂ X − (A).
Proof: (If) Suppose N ⊂ X − (A), which is equivalent
where P ∈ R −1 , R ∈ Rm×k and Im P ⊂ V. Define T =
n×k
⊥
to X − (A) ⊂ N ⊥ , which evaluates to X + (AT ) ⊂ S1 S2 ; then, T = col(Q1 , Q2 ). Also let V1 := VV and
AT | Im CT . Therefore, (AT , CT ) is patterned stabilisable V2 := VW . Then T−1 AS1 = T−1 P − T−1 BR decomposes
by Theorem 5.3, so there exists K ∈ C(Y T , V T ) for which as
σ (AT + CT K ) ⊂ C− . Take L := −K T ∈ C(V, Y ), giving
A1 = P1 − B1 R (11)
−
σ (A − LC) = σ (A − LC) T
= σ (A + C K ) ⊂ C .
T T
A2 = −B2 R (12)
1498 A. C. SNIDERMAN ET AL.
where Ai := Qi AS1 ∈ C(Vi , V1 ), Bi := Qi B ∈ C(Vi , U ) Proof: Consider the recursive algorithm in (Wonham,
and Pi Qi P for i = 1, 2; the commuting relationships 1979, Theorem 4.3)
follow from Lemma 3.4(i) and (ii), and P2 = Q2 P =
0 since Im P ⊂ V. Thus, finding P and R that satisfy V 0 = Ker(D)
Equation (10) is equivalent to finding P1 and R that V i = Ker(D) ∩ A−1 (B + V i−1 ).
satisfy Equations (11) and (12). By Lemma 3.1, V1 and
V2 are diagonalisable, so by Lemma 3.7, Equation (12) The sequence is non-increasing and has a lower bound
can be solved for R ∈ C(U, V1 ). Then, P1 A1 + B1 R, of {0}, so it must have a fixed point. By Wonham (1979,
and P1 ∈ C(V1 ) by Lemma A.1. Since Im P ⊂ V, there- Theorem 4.3), that fixed point is V
. Also, using Lemma
fore S1 Q1 P = P, so P S1 P1 . Now define the friend 3.2 and Lemma 3.3(iii), each V i is a V-decoupling sub-
F = RQ1 . Observe that (A + BF)S1 = AS1 + BRQ1 S1 space. Hence, the fixed point V
is also a V-decoupling
= AS1 + BR = P, so (A + BF )V = Im P ⊂ V. Also, subspace.
F ∈ C(U, V ) by Lemma 3.5(iii). We conclude F is a pat-
terned friend of the V-decoupling controlled invariant The Patterned OSP is solved next, using the same solv-
subspace V. ability condition as the standard, unpatterned OSP.
Theorem 7.2: Let (D, A, B) be a patterned system. The
Given a controlled invariant subspace V with a friend
Patterned OSP is solvable if and only if
F, any system can be transformed into a basis which splits
it into the parts in V and the parts not in V. If the sys-
X + (A) ⊂
A|B + V
(14)
tem, subspace and friend are all patterned, then each of
the transformed parts will also be patterned, based on the
where V
:= sup I(Ker D).
Patterned Representation Theorem 3.1. These commut-
ing relationships are given in the following Third Decom- Proof: (If) Suppose Equation (14) holds. By Lemma 7.2,
position Theorem. The proof follows immediately from V
is a controlled invariant and V-decoupling subspace.
Theorem 3.1 and Lemma 7.1. Let W be a V-invariant complement. Then we can apply
Theorem 7.1 (Third Decomposition Theorem): Let (A, the Third Decomposition Theorem 7.1 with (ξ 1 , ξ 2 ) =
B) be a patterned system, and let V ⊂ X be a V-decoupling T−1 x, v = u − Fx and F ∈ C(U, V ) a patterned friend of
and controlled invariant subspace with V-invariant com- V
, to get
plement W. Then, there exists a state and feedback trans-
formation (T, F), where T : X → X and where F ∈ ξ˙1 Ã1 ∗ ξ1 B
= + 1 v
C(U, V ) is a patterned friend of V, such that the trans- ξ˙2 0 Ã2 ξ2 B2
formed system matrices have the form (15)
such that σ (Ã2 + B2 K2 ) ⊂ C− . Lifting K2 and apply- Problem 9.1 (Patterned SMFP): Let (C, A, B) be a pat-
ing Lemmas 3.5 and A.1, we obtain the overall pat- terned system satisfying Assumption 3.1. Find a pat-
terned feedback K = F + K2 Q2 ∈ C(U, V ). It follows terned state feedback K : X → U, K ∈ C(U, V ), such
from Lemma 2.1 that X + (A + BK ) ⊂ V
. Thus, K solves that
the Patterned OSP.
(Only If): Condition (14) is exactly the necessary con- Ker C ⊂ Ker K (16)
dition for the general OSP. Since it is necessary for the
existence of a general state feedback, it is also necessary σ (A + BK ) ⊂ C− . (17)
for the existence of a patterned state feedback.
Condition (17) guarantees the stability of the closed-
loop system, and condition (16) ‘masks out’ any state
8. Disturbance decoupling
information that is not available in the measurement
Consider the system (D, A, B, E) from Equation (6), where y, which ensures that the state feedback K can later
E corresponds to a disturbance w(t) that is not directly be turned into a measurement feedback K . This mask
measurable by the controller. The Disturbance Decou- is characterised geometrically by L :=
A| Ker C, the
pling Problem (DDP) is to find a state feedback u(t) = smallest A-invariant subspace containing Ker C. Then,
Kx(t) so that the output z(t) is unchanged for any dis- any feedback K found on the subsystem corresponding
turbance w(t). This control problem can be stated in the to L will satisfy L ⊂ Ker K, fulfilling condition (16).
same manner for patterned systems with patterned feed- Lemma 9.1: Let (C, A) be a patterned system satisfy-
back, and is given in geometric terms as follows. ing Assumption 3.1. Then, the subspace L :=
A| Ker C =
Problem 8.1 (Patterned DDP): Let (D, A, B, E) be a Ker C + A Ker C + · · · + An−1 Ker C is a V-decoupling
patterned system satisfying Assumption 3.1. Find a pat- subspace.
terned state feedback K : X → U, K ∈ C(U, V ), such Proof: A ∈ C(V ) and C ∈ C(Y, V ). By Lemma A.1, Ai ∈
that
A + BK| Im E ⊂ Ker D. C(V ) for all i ∈ N, and by Lemma 3.3(i), Ker C is a
V-decoupling subspace; therefore, by Lemma 3.3(iii),
The Patterned DDP and the standard, unpatterned
Ai Ker C is a V-decoupling subspace. Using Lemma 3.2,
DDP are solvable by the same condition, as shown next.
it follows that L is a V-decoupling subspace.
The proof is a minor variation of Wonham (1979, Theo-
rem 4.2), so it is omitted here. With the guarantee that the masking subspace L :=
A| Ker C is a V-decoupling subspace, the Patterned
Theorem 8.1: The Patterned DDP is solvable if and only
SMFP has the following sufficient conditions for the exis-
if E ⊂ V
, where E = Im E and V
= sup I(Ker D).
tence of a solution – the same conditions as for the stan-
The above solution to the Patterned DDP does not dard, unpatterned SMFP.
depend on E being a patterned matrix. It is interesting that Theorem 9.1: The Patterned SMFP is solvable if
a patterned feedback can block out disturbances that do
not follow the system’s pattern and underlying structure.
X + (A) ⊂
A|B (18)
The SMFP asks for a measurement feedback u = K y; Proof: (If) Suppose Equations (23), (24) and (25) hold for
however, the above solution to the Patterned SMFP gave some V-decoupling subspace V ∈ I(Ker D). By Lemma
a state feedback u = Kx. Since Ker C ⊂ Ker K (from 3.2, L + V and L ∩ V are V-decoupling subspaces, and
requirement (16)), the state feedback K can be turned into have V-invariant complements R and W, respectively;
a measurement feedback K that solves the equation K C that is,
= K; also, since K ∈ C(U, V ) and C ∈ C(Y, V ), a pat-
X = (L + V ) ⊕ R, X = (L ∩ V ) ⊕ W.
terned solution K ∈ C(U, Y ) can be found using Lemma
3.7. Thus, the patterned measurement feedback u = K y
Also, L̃ := L ∩ W and Ṽ := V ∩ W are V-decoupling
solves the Patterned SMFP. These comments can also be
subspaces by Lemma 3.2. By the modular distributive rule
carried forward to patterned observers for detectable sys-
(Wonham, 1979, (0.3.1)), L = (L ∩ V ) ⊕ L̃ and V =
tems.
(L ∩ V ) ⊕ Ṽ. Thus, the state space X splits into four V-
decoupling subspaces:
10. Output stabilisation by measurement
X = (L ∩ V ) ⊕ L̃ ⊕ Ṽ ⊕ R. (26)
feedback
Consider the system (D, C, A, B) from Equation (6). The Let Si and Qi , i = 1, … , 4, be the insertion and natural pro-
Output Stabilisation by Measurement Feedback Problem jection maps of these spaces, in the same order as Equa-
(OSMFP) is to find a measurement feedback u(t) = K y(t) tion (26).
such that z(t) → 0 as t → . As in the SMFP, the geomet- Next, a patterned friend F of V will be found such
ric solution to the OSMFP instead finds a state feedback that L and V are both (A + BF )-invariant. We start
u(t) = Kx(t), where Ker K ⊃ L for a ‘masking subspace’ by choosing any patterned friend F ∈ C(U, V ) using
L ⊃ Ker C that blocks out any state information not seen Lemma 7.1, and then take F FS3 Q3 . By Lemmas 3.4(iii)
in the measurement. This turns the OSMFP into the and 3.5(iii), F ∈ C(U, V ). Also, F L = FS3 (Q3 L) = 0
Restricted Regulator Problem (RRP), which can be stated since L ⊂ Ker Q3 . Thus, (A + BF )L = AL ⊂ L, so L is
geometrically for patterned systems in the same way as (A + BF )-invariant. Finally, V is also (A + BF )-invariant:
for unpatterned systems. We show that the Patterned RRP using F L = 0, Ṽ = S3 Q3 Ṽ, and Equation (23),
and the standard, unpatterned RRP are equally solvable;
that is, if any feedback exists, then a patterned feedback (A + BF )V = (A + BF )(L ∩ V ) + (A + BF )Ṽ
exists. = A(L ∩ V ) + (A + BFS3 Q3 )(S3 Q3 Ṽ )
= A(L ∩ V ) + (AS3 + BFS3 )(Q3 Ṽ )
Problem 10.1 (Patterned RRP): Let (D, A, B) be a
patterned system, and let L be an A-invariant and V- = A(L ∩ V ) + (A + BF )(S3 Q3 Ṽ )
decoupling subspace. Find a patterned state feedback K : ⊂ (L ∩ V ) + V = V.
INTERNATIONAL JOURNAL OF CONTROL 1501
Therefore, F ∈ C(U, V ) is a patterned friend of V for V-decoupling and controlled invariant subspace, as
which L, V, L ∩ V and L + V are all (A + BF )-invariant. required.
Now,we apply thefeedback transformation (T, F ), where Since L ⊂ Ker K by Equation (21), (A + BK )L =
T = S1 S2 S3 S4 . Let ξ = T−1 x and v = u − F x. AL . Then using the property X + (A) ∩ L = X + (A +
Then BK ) ∩ L = V ∩ L (Wonham, 1979, Lemma 6.1), con-
⎡ ⎤ dition (24) holds. Second, since L and X + (A) are
Ã1 ∗ ∗ ∗ both A-invariant, condition (23) also holds. Third,
⎢ ⎥ using the property
A|B + X + (A + BK ) =
A|B +
⎢ 0 Ã2 0 ∗ ⎥
ξ˙ = ⎢ ⎥ξ
⎣ 0 0 Ã3 ∗ ⎦ X + (A) (Wonham, 1979, Lemma 6.2), we have X + (A) ⊂
0 0 0 Ã4
A|B + X + (A) =
A|B + V by definition of V, so con-
⎡ ⎤ dition (25) holds.
B1
⎢ B2 ⎥
+⎢ ⎥
⎣ B3 ⎦ v (27)
The Patterned OSMFP asks for a measurement feed-
B4 back u = K y; however, in converting the Patterned
OSMFP to the Patterned RRP, the above solution pro-
where the zeros arise from the (A + BF )-invariance of L, vided a state feedback u = Kx. In that solution, K was
V, L ∩ V and L + V. Also, all the blocks are patterned as ‘restricted’ such that Ker K ⊃ L for some V-decoupling
in Lemmas 3.4(i) and (ii); in particular, Ã4 ∈ C(VR ) and subspace L. If also L ⊃ Ker C, then the patterned state
B4 ∈ C(VR , U ). feedback K can be transformed into a patterned mea-
We claim that (Ã4 , B4 ) is stabilisable. Using Equation surement feedback K ∈ C(U, Y ) by solving the equation
(25) and the property X + (A + BF ) ⊂
A + BF |B + V K C = K, using Lemma 3.7. This operation depends on
(Wonham, 1979, Lemma 6.2), the choice of the V-decoupling subspace L; one possi-
ble choice is L =
A| Ker C, as in the SMFP. Then, the
R+ (Ã4 ) = Q4 X + (A + BF ) ⊂ Q4
A + BF |B + Q4 V patterned measurement feedback u = K y solves the Pat-
=
Ã4 |B4 . terned OSMFP.
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
1 0 1 0 2 0 1 0 0 0
⎢ 1 0 −1 0 ⎥ ⎢ −3 0 0 1⎥ ⎢1 0⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ 0 1 0 1⎥ ⎢ 0 2 1 0⎥ ⎢0 0⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ 0 1 0 −1 ⎥ ⎢ 0 −3 0 −1 ⎥ ⎢0 1⎥
⎢ ⎥ N = Im ⎢ ⎥, V = Im ⎢
⎥.
B=⎢ ⎥, ⎢ −2 0⎥ ⎢0 0⎥
⎢ 1 0 1 0⎥ ⎢ 0 1 ⎥ ⎢ ⎥
⎢ ⎥ ⎢ 3 1⎥ ⎢1 0⎥
⎢ −1 0 1 0 ⎥ ⎢ 0 0 ⎥ ⎢ ⎥
⎢ ⎥ ⎣ 0 −2 1 0⎦ ⎣0 0⎦
⎣ 0 1 0 1⎦
0 −1 0 1 0 3 0 −1 0 1
⎡ ⎤
7 5 −1 1 −5 −3 −1 1
⎢ −1 1 7 5 −1 1 −5 −3 ⎥ The columns of each matrix form a basis for the corre-
C=⎢
⎣ −5 −3 −1 1 7 5 −1 1 ⎦,
⎥
sponding subspace. Denote the ith basis vector of C by ci ,
−1 1 −5 −3 −1 1 7 5 of N by ni , of V
by vi and of X + (A) by ai .
⎡ ⎤ If the system were fully stabilisable and detectable, a
5 0 −1 0 1 0 −1 0
patterned feedback could be found using the techniques
⎢ −1 0 5 0 −1 0 1 0 ⎥
D=⎢ ⎥
⎣ 1 0 −1 0 5 0 −1 0 ⎦.
of the Patterned SMFP; this would stabilise the full sys-
tem, including the output z. However, the unstable mode
−1 0 1 0 −1 0 50 a3 is not controllable, so the full system cannot be sta-
bilised by any feedback, patterned or otherwise. Instead,
since a3 ∈ V
, this uncontrollable unstable mode does
not show up in the output z, so the output can be sta-
The system’s base matrices are determined by the bilised by turning the OSMFP into a Patterned RRP, as
fundamental permutation matrix 4 , as in Example in Section 10. Formulating the Patterned RRP requires
3.1 – specifically, A ∈ C(4 ⊗ I2 ), B ∈ C(4 ⊗ I2 , 4 ⊗ a choice of two (4 I2 )-decoupling subspaces: we take
I) and C, D ∈ C(4 ⊗ I, 4 ⊗ I2 ) (where ‘’ denotes the L :=
A| Ker C = N and V := V
. Using these sub-
Kronecker product). The control objective is to stabilise spaces, we verify the solvability conditions (23)–(25) of
the system output z = Dx using a patterned measurement the Patterned RRP, given in Theorem 10.1: first, L ∩ V =
feedback u = K̃y with K̃ ∈ C(4 ⊗ I) – this is the Pat- span{n4 } is A-invariant, satisfying Equation (23); sec-
terned OSMFP, as in Section 10. ond, X + (A) ∩ L = span{n4 } = L ∩ V, satisfying Equa-
The eigenvalues of A are given by tion (24) and third,
A|B + V = X ⊃ X + (A), satisfying
the (4 I2 )-patterned spectrum σ (A) = Equation (25), and confirming that the Patterned RRP is
{−16, −12, −4, −4, −4, 8, 8, 8}. The controllable solvable.
subspace C :=
A|B, unstable subspace X + (A) := Following the method of Theorem 10.1, a patterned
Ker(A − 8I), unobservable subspace N and supremal output-stabilising feedback can be found by placing
controlled invariant subspace V
:= sup I(Ker D) of the poles in the subsystem corresponding to the (4 I2 )-
system are all (4 I2 )-decoupling, and given by decoupling subspace R := (L + V )⊥ = span{r1 , r2 , r3 },
⎡ ⎤ where r1 = (1, 0, −1, 0, 1, 0, −1, 0), r2 = (3, 2, 0, 0, −3,
1 0 0 0 0 0
⎢0 0 0 0 1 0⎥ −2, 0, 0) and r3 = (0, 0, 3, 2, 0, 0, −3, −2). R is the fourth
⎢ ⎥
⎢0 1 0 0 0 0⎥ subspace in the state space decomposition (26), which has
⎢ ⎥
⎢0 0 0 0 0 1 ⎥ insertion maps
C = Im ⎢ ⎢0 0 1 0
⎥,
⎢ 0 0⎥⎥
⎢ 0 0 0 0 −1 0⎥
⎢ ⎥ S1 = [ n4 ], S2 = [ n1 n2 n3 ],
⎣0 0 0 1 0 0⎦ S3 = [ v ], S4 = [ r1 r2 r3 ],
0 0 0 0 0 −1
⎡ ⎤
1 0 0
⎢ 0 0 1⎥ where v v1 + v2 . Define the coordinate transfor-
⎢ ⎥ mation T = [ S1 S2 S3 S4 ]; then, the natural projection
⎢ 0 1 0⎥
⎢ ⎥ maps Q1 , Q2 , Q3 and Q4 are uniquely defined by T −1 =
⎢ 0 0 −1 ⎥
X + (A) = Im ⎢
⎢ −1
⎥, col(Q1 , Q2 , Q3 , Q4 ), where Q1 is the first row, Q2 is the
⎢ 0 0⎥⎥
⎢ 0 0 1⎥ following three rows, Q3 is the fifth row and Q4 is the
⎢ ⎥
⎣ 0 −1 0⎦ final three rows. With these projection maps, the system
0 0 −1 decomposition can be carried out, beginning by finding a
patterned friend F of V. Following Lemma 7.1, a friend is
INTERNATIONAL JOURNAL OF CONTROL 1503
400 80
System Outputs
System States
200
40
0
−200 0
−400
−40
0 0.1 0.2 0.3 0.4 0 0.5 1 1.5
Time [s] Time [s]
Hazewinkel, M., & Martin, C.F. (1983). Symmetric linear sys- notational conventions from Section 3. First, we require
tems: An application of algebraic systems theory. Interna- the following basic properties of commuting matrices.
tional Journal of Control, 37(6), 1371–1384.
Holmes, C., & Broucke, M.E. (2016). Pattern preserving pole Lemma A.1: The following algebraic properties hold:
placement and stabilization for linear systems. In Proceed-
ings of the American Control Conference. (pp. 5181–-5186). (i) If A, B ∈ C(V, U ), then A + B ∈ C(V, U ).
Boston, MA: IEEE (ii) If A ∈ C(W, V ) and B ∈ C(V, U ), then AB ∈
Hovd, M., Braatz, R.D., & Skogestad, S. (1997). SVD controllers
for H2-, H- and μ-optimal control. Automatica, 33(3),
C(W, U ).
433–439. (iii) If A ∈ C(V, U ), then AT ∈ C(U T , V T ).
Hovd, M., & Skogestad, S. (1994). Control of symmetrically (iv) If A ∈ C(V, U ) is non-singular, then A−1 ∈
interconnected plants. Automatica, 30(6), 957–973. C(U, V ).
Lin, C.T. (1974). Structural controllability. IEEE Transactions
on Automatic Control, 19(3), 201–208.
Lunze, J. (1986). Dynamics of strongly coupled symmetric Let the Jordan forms of V and U be given by
composite systems. International Journal of Control, 44(6),
1617–1640. V̂ := V−1V V = diag(V̂1 , . . . , V̂ν ),
Lunze, J. (1989). Stability analysis of large-scale systems com-
Û := U−1U U = diag(Û1 , . . . , Ûμ )
posed of strongly coupled similar subsystems. Automatica,
25(4), 561–570.
Massioni, P., & Verhaegen, M. (2009). Distributed control for where each V̂i and Û j is a Jordan block. Block diagonal-
identical dynamically coupled systems: A decomposition isation of commuting matrices is based on these Jordan
approach. IEEE Transactions on Automatic Control, 54(1), blocks.
124–135.
Rotkowitz, M., & Lall, S. (2006). A characterization of convex Theorem A.1 (Gantmacher, 1959): Let A and  satisfy
problems in decentralized control. IEEE Transactions on  := V−1 AU , and partition  according to the Jordan
Automatic Control, 51(2), 1984–1996.
blocks of V and U:
Sniderman, A.C., Broucke, M.E., & D’Eleuterio, G.M.T. (2013).
Block circulant control: A geometric approach. In Proceed- ⎡ ⎤
ings of the 52nd IEEE Conference on Decision and Control Â11 ··· Â1μ
(pp. 2097–2102). Florence, Italy: IEEE. −1 ⎢ .. .. ⎥
 := V AU = ⎣ . . ⎦
Sniderman, A.C., Broucke, M.E., & D’Eleuterio, G.M.T.
(2015a). Controllability is not sufficient for pole placement Âν1 · · · Âνμ
in patterned systems. In Proceedings of the 54th IEEE Con-
ference on Decision and Control (pp. 4296–-4301). Osaka,
Japan: IEEE. where Âi j has the same number of rows as V̂i and the same
Sniderman, A.C., Broucke, M.E., & D’Eleuterio, G.M.T. number of columns as Û j . Also suppose that V̂i corresponds
(2015b). Formation control of balloons: A block circulant to eigenvalue δ i , and Û j corresponds to eigenvalue γ j . Then,
approach. In Proceedings of the American Control Confer-
ence (pp. 1463–1468). Chicago, IL: IEEE. A ∈ C(V, U ) if and only if the blocks Âi j satisfy the follow-
Sundareshan, M.K., & Elbanna, R.M. (1991). Qualitative analy- ing for all i, j: if δ i = γ j , then Âi j is an upper triangular
sis and decentralized controller synthesis for a class of large- Toeplitz matrix; and if δ i γ j , then Âi j = 0.
scale systems with symmetrically interconnected subsys-
tems. Automatica, 27(2), 383–388. Any matrix A ∈ C(V, U ) (for any V and U) can be put
van der Schaft, A. (1987). Symmetries in optimal control. SIAM
into the partitioned form of  in Theorem A.1. Combin-
Journal on Control and Optimization, 25(2), 245–259.
Wang, S.H., & Davison, E.J. (1973). On the stabilization of ing this result with Assumption 3.1 that V and U are diag-
decentralized control systems. IEEE Transactions on Auto- onalisable, each upper triangular Toeplitz matrix Âi j is a
matic Control, 34(4), 473–478. scalar because the Jordan blocks of V and U are all 1 × 1.
Wonham, W.M. (1979). Linear multivariable control: A geomet- Thus, in our case, a pattern C(V, U ) determines a ‘hatted’
ric approach (3rd ed. ). New York, NY: Springer-Verlag.
form in which some entries are fixed at zero, while others
can be freely assigned. Furthermore, the partitioning of
Theorem A.1 results in a block diagonal matrix if V and
Appendix. Block diagonalising patterned matrices U are chosen according to the ordering conventions of
There are two places in our framework where pat- Remark 3.1. The conventions regarding V and U also
terned matrices are explicitly synthesised: in solutions carry over to their inverses.
to linear equations (Lemma 3.7), and in pole placement Lemma A.2: Suppose V and U are real matrices, and fol-
(Theorem 5.1). This appendix provides further back- low Remark 3.1. Partition V = row( 1 , . . . , r ) as in
ground on commuting matrices and their block diago- (4), and partition V−1 as col( 1 , . . . , r ), where the num-
nalisation in order to complete those proofs. Recall the ber of rows of i equals the number of columns of i . Let
1506 A. C. SNIDERMAN ET AL.
¯ ε for each i.
i
¯ i
A.2, and Âi = Âε by assumption. There are two cases: if
Proof: Since V−1 V = I, therefore i i = I, and i j = εi = i, then i Âi i is real; and if εi = j i, then i Âi i =
¯ ε ¯ εi .
i
0 for all i j. Also, ε εi = I¯ = I, and so i i = j  j j and so i Âi i + j  j j is real. We conclude A
i
Lemma 3.6. ¯ i ¯i
Proof of Lemma 3.6: (If) Let A ∈ CR (V, U ). By the Âi = Âε and B̂i = B̂ε . Therefore, solving the equation
ordering of Remark 3.1 and by Theorem A.1, Â = XA = B for X is equivalent to solving the equation
diag(Â1 , . . . , Âr ). (Note that the Âi blocks may con-
⎡ ⎤⎡ 1 ⎤ ⎡ 1 ⎤
tain multiple Â(··) blocks of Theorem A.1). Partition X̂ 11 · · · X̂ 1r  B̂
V−1 = col( 1 , . . . , r ) as in Lemma A.2 and V = ⎢ .. .. ⎥ ⎢ .. ⎥ ⎢ .. ⎥
⎣ . . ⎦⎣ . ⎦=⎣ . ⎦ (A2)
row( 1 , . . . , r ) as in Equation (4), where i has ni
columns and i has mi rows. Then, X̂ r1 · · · X̂ rr Âr B̂r
⎡ ⎤ ⎡ ⎤
Â1 1 A 1
for X̂ i j ∈ C p ×n . This decomposes into
i j
⎢ .. ⎥ ⎢ .. ⎥
 := ⎣ . ⎦=⎣ . ⎦.
 r r A r
X̂ i j  j = 0, i = j
¯ ε , and by Remark 3.1, i = ¯ εi .
i
X̂ ii Âi = B̂i , i = 1, . . . , r.
By Lemma A.2, i =
¯ i
Since A is real, it follows that Âi = Âε , for i = 1, … , r.
(Only If): Let  = diag(Â1 , . . . , Âr ), where Âi ∈ For the first set of equations, we choose the solution X̂ i j =
¯ 0, i j. For the second set of equations, we choose any
Cn ×m and Âε = Âi for each i = 1, … , r. Define A :=
i i i
¯ ¯ ¯ ¯
solution such that X̂ ε ε = X̂ ii (note that X̂ ii Âε = X̂ ii Âi =
i i i
−1
V ÂU . Then we have
¯
X̂ ii Âi = B̂i = B̂ε ).
i
each i = 1, … , r. ¯ i
chosen such that K̂ i = K̂ ε . Now, define the overall feed-
Define  := V−1 AV and B̂ := V−1 BU . By Lemma
back K := V diag(K̂ 1 , . . . , K̂ r )U−1 . Then, K ∈ C(V, U )
3.6, Â = diag(Â1 , . . . , Âr ) and B̂ = diag(B̂1 , . . . , B̂r ),
by Lemma 3.6, and σ (A + BK ) = L by construction, so
where Âi ∈ Cn ×n and B̂i ∈ Cn ×m . Since (A, B) is con-
i i i i