LATEX template for Hindawi journal articles

1 Journal of Mathematics

2 Research Article

3 The category of linear systems is object semisimple

4
M.V. Carriegos1
1
Departamento de Matemáticas. Universidad de León. León, Spain.

Correspondence should be addressed to M.V. Carriegos: miguel.carriegos@unileon.es

Emails: miguel.carriegos@unileon.es (M.V. Carriegos)

5 Abstract
6

7 Keywords: keyword; keyword; keyword

8 Introduction
9 Mathematical study of control systems arises in 19th century from engineering after seminal work of J.C. Maxwell
10 [1]. Linear systems arise everywhere in the topic [2] both as linear models and as linealizations. Algebraic study
11 of linear systems ([3], [4]) deals with linear systems defined on algebras and modules over a commutative ring, at
12 least in the state-space approach [5]. This approach has been used recently in the field of convolutional coding [6],
13 [7], [8], [9], [10], [11], [12]; and in the field of cybersecurity (Kuijper, Maite).
14 Feedback is the main tool in the state-space approach [13], and two linear systems are feedback equivalent if
15 one can be transformed in the another by means of a feedback action. This equivalence relation has been studied
16 for decades from many different perspectives (citar aqui). We would like to remark two concrete classical results
17 we are using as cornerstones in this paper. The Canonical Controller Form [2], and Brunovsky’s Theorem [14].
18 The detection of feedback invariants as well as the introduction of notion of feedback morphism [15] as a linear
19 map preserving both the dynamics and the controls of linear systems, was introduced to circumscribe the problem
20 of feedback classification of linear systems over R-modules. The notion of feedback morphism is interesting by
21 itself in terms of linear systems over K-vector spaces [16].
22 Categories of linear systems arise [17], [18]
23 This paper deals with the study of exact structures on the additive categories of linear systems. Definition of
24 (categorical) subsystem of a linear system is given and hence simple linear systems arise as those nonzero linear
25 systems with no nontrivial subsystems. Reachable simple systems are characterized in this paper as those systems
26 isomorphic to a Canonical Controller Form. Brunovky’s Theorem implies that the category of reachable linear
27 systems is object semisimple. But the category itself fails to be semisimple because it is not abelian and Schur’s
28 Lemma does not hold in this category (see [19]).

1
 The exact structure of feedback actions

1
29 The paper is structured as follows: Main definitions (linear system, feedback morphism, reachable system,
30 and reachability map) are found in the second section of Preliminaries; the categories SK of linear systems, and
31 AK of reachable systems are also reviewed as well as some properties of feedback morphisms. Third section is
32 the main section of the paper, it is devoted to prove that categories SK and AK are pre-abelian, i.e. they have all
33 kernels and all cokernels of (feedback) morphisms; proofs are constuctive, we compute effectively both kernels
34 and cokernels , images and coimages of feedback morphisms in SK and in AK . Fourth section deals with main
35 decomposition theorems of linear systems; we state Kalman’s Decomposition Theorem as a split exact sequence in
36 SK , and Brunovsky’s Classification Theorem in terms of short exact sequences in AK . Fifth section is devorted to
37 find out simple objects in category of linear systems as well as to prove semisimplicity of the category. The paper
38 concludes with a final section where some future work is pointed out.

39 Preliminaries
40 A linear system is a triple σ = (V, f, B) where V is a finite-dimensional K-vector space, f : V → V is a linear
41 map, and B ≤ V is a vector subspace (we will use ≤ to denote vector subspace).

42 Feedback morphisms
43 The category SK of linear systems over finite dimensional K-vector spaces gathers linear systems σ = (V, f, B) as
44 objects in the category, and feedback morphisms a : (V, f, B) → (V 0 , f 0 , B 0 ) as morphisms in the category.
45 Recall that, [15, Definition 3.2.], a feedback morphism a : σ = (V, f, B) → (V 0 , f 0 , B 0 ) = σ 0 is given by a
46 linear map F (a) = a : V → V 0 satisfying the following properties

47 (i) a(B) ⊆ B 0

48 (ii) im (f 0 ◦ a − a ◦ f ) ⊂ B 0 .

49 The pair SK = (Linear systems, Feedback morphisms) is a category. In fact a K-linear category (i.e. enriched
50 on the category VK of K-vector spaces and linear maps). The functor forget-the-dynamics F : SK → VK given
51 by F (V, f, B) = V in objects, and by F (a) = a in morphisms is obviously injective on morphisms hence F is a
52 faithful functor. F is also a dense functor because every vector space V occurs as V = F (V, 0, 0). But functor F
53 is not full because not every linear map arises as a feedback morphism, i.e. the induced map

Fσ,σ0 : homSK (σ, σ 0 ) → homK (F (σ), F (σ 0 )) (1)

54 is injective but it is not surjective in general.

55 A key property of feedback morphisms is that they generalize the notion of feedback equivalence [15, Proposi-
56 tion 3.3.]. In fact two linear systems σ and σ 0 are feedback equivalent (in the classical sense) exactly when they are
1 Linear systems [4], [3] are motivated by their applications to control engineering (see for instance [2]), and to convolutional coding (see

[22]). Feedback actions are a key tool to automatic control of a system [13]. A huge research has been performed for decades in several
disciplines dealing with linear control systems. We are interested in the mathematical perspective of feedback actions and feedback equivalences.
Feedback equivalence was characterized [14] in case of linear systems over finite dimensional vector spaces. Feedback actions, feedback
invariants, canonical forms as well as compositions and decompositions of linear systems have been used intensively in applications [6], [7],
[8], [9], [10], [11], [12].
A natural generalization of linear systems over vector spaces to linear systems over R-modules, R being a commutative ring [4], arises
when dealing with convolutional codes over finite rings [6], or dealing with linear control systems depending on parameters [4]. But feedback
classification was shown to be wild over a rather general commutative ring [23].
We prove that the category is pre-abelian by computing effectively kernel and cokernel of any feedback morphism. We also show that the
linear systems categories fail to be semi-abelian (and hence abelian) by setting up concrete examples. Hence categories of linear systems are
located strictly in the pre-abelian stratum of additive categories hierarchy [20, 2.11].
Kernel-Cokernel pairs on the category of linear systems are considered. The minimal exact structure is split ker-coker sequences [21]. It
is sufficient to rewrite classical Kalman’s Decomposition and Brunovsky’s Classification Theorems in terms of split exact sequences in the
pre-abelian category of linear systems.

2
 The exact structure of feedback actions

57 isomorphic in SK , i.e. when there exists a feedback mophism a : σ → σ 0 such that its inverse a−1 : σ 0 → σ is also
58 a feedback morphism.

59 Reachable systems
60 Consider a linear system σ = (V, f, B). Consider the subspaces of V given by: N0σ = 0, and recursively Nkσ =
σ


61 B + f Nk−1 . This sequence of subspaces is an ascending chain (i.e. Nkσ ≤ Nk+1
σ
). The chain is strict up to an
62 index (the degree of the linear system) s = deg(σ) ≤ dim V , and from this index on the chain estabilizes forever

N0 < N1 < · · · < Ns−1 < Ns = Ns+1 = · · · (2)

63 Since V is finite dimensional then Cayley-Hamilton Theorem implies in particular that deg(V, f, B) ≤ dim V . We
64 will often use the notation f ∗ (B) to denote f ∗ (B) = B + f (B) + f 2 (B) + · · · + f n (B) + · · · ≤ V .
σ ∗
65 Linear system σ = (V, f, B) is called reachable if Ndeg σ = f (B) = V . Denote by AK the full subcategory
66 of SK collecting all reachable systems and all feedback morphisms in SK between reachable systems.
67 Consider the restriction of forget-the-dynamics functor F to reachable systems G : AK → VK . Functor G is
68 newly injective on morphisms hence G is faithful. Functor G is also dense because every vector space V occurs as
69 V = G(V, 0, V ), and (V, 0, V ) is trivially a reachable linear system. But G is not full because the induced map

Gσ,σ0 : homAK (σ, σ 0 ) → homK (G(σ), G(σ 0 )) (3)

70 is not surjective in general.

71

72 A remarkable class of reachable linear systems is the class of Canonical Controller Forms. The Canonical
73 Controller Form (CCF) over Kn is the following reachable linear system

0 ... ... ... 0

     
1
.. .
. . 0
1 .
   * +
    .  
n  n  .. .. . .
Σ = K ,  ,  (4)
.
  .  
0 . . .  
    .  
. .. .. .. . .
. . .
   
. . . . .
0
0 ... 0 1 0

74 Following classical result in control theory shows that feedback classification is trivial for so-called single-input
75 reachable systems.

76 Theorem 1 (Canonical Controller Form) Let σ = (V, f, B) be a reachable linear system such that dim B = 1
77 then σ ∼
= Σdim V

78 Proof. Set n = dim V and B =< ~b >= b. Since σ is reachable it follows that V = f ∗ (B) = b + f b + · · · + f n−1 b.


79 Now consider the basis of V given by P = b, f b, · · · , f n−1 b . Then the feedback isomorphism given by selecting
80 basis P is the following
  −a1 −a2 ··· −an−1 −an
  1
 
1 0 ··· 0 0
   * +
ϕP .. . .
. .   = σ0
 n     
σ = (V, f, B) −−→ K , 

 0 1 . . , 
  .
  (5)
. .. .. .. . .
  . . . . .   .  
. . .
0 ··· 0 1 0 0

81 And it is straightforward to check that σ 0 ∼

= Σn .
82 

3
 The exact structure of feedback actions

83 Monics, epics, sections, and retracts

a
84 A feedback morphism σ −−→ σ 0 is a monomorphism (or a is monic) if a ◦ f = a ◦ g for feedback morphisms
85 f, g : σ00 → σ implies f = g. Since F : SK → VK is faithful then F reflects monomorphisms, i.e. if F (a) is monic
86 (i.e. a injective linear map) then a is monic. The converse result does not hold (see [18, Ex. 5.2]), i.e. monics are
87 not preserved in general by F .
a
88 On the other hand, σ −−→ σ 0 is a epimorphism (or a is epic) if f ◦ a = g ◦ a for feedback morphisms
00
89 f, g : σ → σ implies f = g. Since F : SK → VK is faithful then F reflects epimorphisms, i.e. If F (a) is epic then
90 a is epic. The converse is also true in SK (see [18, Th. 4.1] ), i.e. F preserves epimorphisms. Therefore feedback
91 morphism a is epic if and only if linear map F (a) is epic, i.e. if F (a) is onto.
92 It is easy to see that same results apply on category AK i.e. funtor F reflects both monomorphisms and epimor-
93 phisms of reachable systems, but F only preserves epimorphisms of reachable systems.
94 Category SK is not balanced, in fact (see [18, Th. 4.3]) identity matrix (1) ∈ K1×1 yields a bimorphism
(1)
95 (K, 0, 0) −−→ (K, 0, K) that is not an isomorphism. Category AK is also not balanced because the morphism
f
   D E (1,0)
0 0 1
Σ2 = K 2 , 1 0
, 0
−−−→ (K, 0, K) = Σ1 (6)
g

96 is a bimorphism of reachable linear systems that is not an isomorphism.

97 Feedback morphism a : σ → σ 0 is a section if it is left invertible i.e. there exists a0 : σ 0 → σ such that
0
98 a ◦ a = 1σ . Every functor preserves sections thus if feedback morphism a is a section then linear map F (a) is a
99 section (i.e. injective). But sections are not reflected by F , the feedback morphism [18, 5.2]
 
1
0
   
0 1
(K, 0, 0) −−−−→ K2 , 0 0
,0 (7)
h

100 is not a section in SK but however F (h) is a injective linear map and hence h is monic but not a section.
101 Morphism a : σ → σ 0 is a retract if it is right invertible i.e. there exists a0 : σ → σ 0 such that a ◦ a0 = 1σ0 .
102 Every functor retracts thus if a is a retract in SK then F (a) is a surjective linear map. But retracts are not reflected
103 by forget functor F : above morphism g that is neither a retract in AK nor in SK though F (a) = (1, 0) is a surjective
104 linear map.

105 Category of linear systems is pre-abelian

106 Categories of linear systems are additive: the parallel gathering of systems (V, f, B) and (W, g, D) given by
   
f 0
(V, f, B) ⊕ (W, g, D) = V ⊕ W, f ⊕ g = 0 g
,B ⊕ D (8)

107 is biproduct (both product and coproduct) in SK [15, Lemma 3.5.]. Hence the category is additive, in fact symmetric
108 monoidal. The zero object is linear system 0 = (0, 0, 0). If linear systems σ, σ 0 are reachable then system σ ⊕ σ 0 is
109 also reachable. On the other hand zero system (0, 0, 0) is reachable of degree zero. Hence ⊕ is internal to AK and
110 reachable systems is also an additive (in fact symmetric monoidal) category.
111 A pre-abelian category is an additive category having all kernels and all cokernels. This section is devoted to
112 compute cokernels and kernels in SK and in AK . As a consequence we will have that both categories SK and AK
113 are complete (kernels, cokernels, limits, and colimits). Our proofs will be constructive: we compute effectively
114 both the cokernel and the kernel of a morphism in SK and in AK . Note that a shorter purely existencial proof might
115 be performed.

4
 The exact structure of feedback actions

116 The cokernel of a feedback morphism

117 The cokernel of a feedback morphism a : σ → σ 0 is a pair Coker(a) = (Coker(a), c) where Coker(a) is a linear
118 system and c : σ 0 → C is a morphism making commutative the left triangle of bellow diagram such that for every
119 linear system σ 00 and every morphism b : σ 0 → σ 00 making right triangle commutative, there exists an unique
120 feedback morpphism f making all diagram commutative

a
0
 0

σ0 ?
 ??
 ??
c ??b
??
 ??
 ??
  
Coker(a) _ _ _ _ _ _ _ _ _/ σ 00
∃!
(9)
f

121 Above diagram is cokernel diagram of feedback morphisms.

122 Cokernels of feedback morphisms in SK and in AK are computed below. First we need to set some technical
123 issues.

124 Definition 2 (First Reduced Form of a feedback morphism) A feedback morphism a is in First Reduced Form
125 (FRF) if it is on the form
 
η 0
0 0
/ V10 ⊕ V20 ,
0 0
b01
         
f11 f12 b1 f11 f12
V1 ⊕ V2 , , im 0 , im b02
f21 f22 b2 a 0 f22
(10)
∼
= 0 0
126 → V10 is a isomorphism, b1 : Kd → V1 , b2 : Kd → V2 , b01 : Kd → V10 , and b2 : Kd → V20 are linear
where η : V1 −
127 maps.

128 Lemma 3 Let a : σ = (V, f, B) → (V 0 , f 0 , B 0 ) = σ 0 be a feedback morphism. Then there exists isomorphisms of
129 linear systems such that the following diagram is commutative

∼
= / coim a ⊕ ker a,
    
f11 f12 b1
σ = (V, f, B) f21 f22
, im b2

 
η 0
a 0 0

 
∼
= / im a ⊕ coker a,
0 0
b01
    
0 0 0 0 f11 f12
σ = (V , f , B ) 0
f21 0
f22
, im b02
(11)

130 Proof. Just consider the Smith normal form of linear map a = F (a) : V → V 0 , i.e. split V = ∼ coim a ⊕ ker a and
0 ∼ 0 dim B 0
131 V = im a ⊕ coker a. Set linear maps β : K dim B
→ V ≥ B and β : K → V ≥ B 0 such that B = im β
0
0 0
132 and B = im β . Then the result follows setting the natural isomorphism η : coim a → im a induced by a; and
0
133 morphisms fij , fij , bi , and b0i defined accordingly. That is to say: f11 = πcoim a ◦f ◦ιcoim a , f12 = πcoim a ◦f ◦ιker a ,
0
134 f21 = πker a ◦ f ◦ ιcoim a , f22 = πker a ◦ f ◦ ιker a , b1 = πcoim a ◦ β, b2 = πker a ◦ β; and f11 = πim a ◦ f 0 ◦ ιim a ,
135 f12 = πim a ◦ f ◦ ιcoker a , f21 = πcoker a ◦ f ◦ ιim a , f22 = πcoker a ◦ f ◦ ιcoker a , b1 = πim a ◦ β 0 , b02 = πcoker a ◦ β 0
0 0 0 0 0 0 0

136 

5
 The exact structure of feedback actions

137 Lemma 4 The following triangle of feedback morphisms is commutative.

0 0
b01
    
f11 f12
im a ⊕ coker a, 0
f21 0
f22
, im b02
5
jjj
jjjj
 
η 0
0 0 jjjj
j
jj
jj jjj
j j
jjj
jjj
  j    
f11 f12 b 1 0
coim a ⊕ ker a, f21 f22
, im b12 0 1
TTT
TTT
TTT
TTT
TTT
TTT
TTT
TTT

η 0
TTT
0 0 T) 
0 0
b01
    
f11 f12
im a ⊕ coker a, 0 0
f22
, im b02
(12)
 
1 0
138 Proof. It is sufficient to prove that 0 1
is a feedback morphism. In other words, to check that

0 0 0 0
b01
     
f11 f12 f11 f12
im 0
f21 0
f22
− 0 0
f22
⊂ im b02
(13)
 
η 0
139 But this follows straightforward from the fact that 0 0
is a feedback morphism.
140 

141 Corollary 5 Let a : σ → σ 0 be a feedback morphism, then there exist feedback isomorphisms such that ϕ0 ◦ a ◦ ϕ
142 is a FRF.

143 Theorem 6 SK has all cokernels.

144 Proof. It is sufficient to compute the cokernel of a FRF feedback morphism. Consider the FRF
 
η 0
0 0
/ im a ⊕ coker a,
00 0
b01
         
f11 f12 b1 f11 f12
coim a ⊕ ker a, , im 0 , im b02
(14)
f21 f22 b2 a 0 f22

0
145 Set the linear system (coker a, (f22 ), im (b02 )). The following is a cokernel diagram:
    
f11 f12 b1
coim a ⊕ ker a, f21 f22
, im b2

 
η 0
0 a 0 0 0


0 00
   0 

f12 b1 f11
im a ⊕ coker a, 0
f22
, im b02 0
s K K
ss KK
ss KK
( 0 1 )sss KK( 0 y )
s KK
ss KK
s KK
sss KK
ss KK
ss KK
 ys %
0 0
(coker a, (f22 ) , im (b2 )) _ _ _ _ _ _ _
∃! _ _ _ _ _ _ _ _ _/ (W, g, D)
(y)
(15)
146 Therefore Coker(a) = (coker(a), (πcoker a ◦ f 0 ◦ ιcoker a ) , πcoker a (B 0 ))
147 

6
 The exact structure of feedback actions

148 Theorem 7 AK has all cokernels

149 Proof.
150 It is sufficient to prove that if both σ and σ 0 are reachable then the cokernel system (coker 0
, im (b02 )) is
 a,(f22 ) 
f 00 f0 b0
151 also reachable. But this is clear: since linear system σ 0 = im a ⊕ coker a, 11
0
12
0
f22
, im b10 is reach-
2
152 able, it follows that
 00 0
∗  0   
f11 f12 b1 ... ... ... ··· ...
0 f 0
b0 = im b0 0
f b 0 0 2
(f ) b 0 0
· · · (f )n 0 −1 0
b
= im a ⊕ coker a (16)
22 2 2 22 2 22 2 22 2

0 ∗
153 Hence (f22 ) (b02 ) im ( b02 f22
0
b02 0 2 0
(f22 ) b2 0 n0 −1 0
· · · (f22 ) b2 ) = coker a, and cokernel system is reachable.
154 
155 Above results yield that if system σ 0 is reachable then the cokernel system of feedback morphism a : σ → σ 0 is
156 again reachable no matter if σ is reachable or not. In other words,

157 Corollary 8 Let a : σ → σ 0 be a feedback morphism. If system σ 0 is reachable then the cokernel system is also
158 reachable.

159 Let a : σ → σ 0 be a feedback morphism of reachable systems. That is, a ∈ homAK (σ, σ 0 ) = homSK (σ, σ 0 ).
160 Note that the cokernel computed in AK equals the cokernel computed in SK , hence the following result holds.

161 Corollary 9 Inclusion functor I : AK → SK sending I(σ) = σ and I(a) = a is right exact (i.e. functor I is
162 additive and preserves cokernels).

163 The kernel of a feedback morphism

164 The kernel of a feedback morphism a : σ → σ 0 is a pair Ker(a) = (Ker(a), i) where Ker(a) is a linear system
165 and i : Ker(a) → σ is a feedback morphism making commutative the right triangle of below diagram such that for
166 every linear system σ 00 and every morphism b : σ 0 → σ 00 making left triangle commutative, there exists an unique
167 feedback morphism f making all diagram commutative

σ 00 ?_ _ _ _ _ _ _ _ _/ Ker(a)
∃!
?? f 
?? 
??b i 
?? 
?? 
?? 
 
σ
0 0

a
& 0 x
σ (17)

168 Above diagram is a kernel diagram of feedback morphisms.

169 This section is devoted to compute kernels of feedback morphisms. General procedure to compute kernels
170 of feedback morphisms are given in bellow Theorems 13 and 15. These Theorems are stated in terms of a so-
171 called Second Reduced Form of a feedback morphism (see Lemma 12 bellow). The Second Reduced Form will be
172 obtained by reducing a FRF using Smith decomposition of linear map b1 . We think it is interesting to remark two
173 particular cases of computation of kernels before setting the general result.

174 Lemma 10 Let a : σ = (V, f, B) → (V 0 , f 0 , B 0 ) be a feedback morphism, let a = F (a) and suppose that ker(a)
175 is f -invariant, i.e. f (ker a) ≤ ker a. Then Ker(a) = (ker a, f, (B ∩ ker a))

7
 The exact structure of feedback actions

176 Proof. The following is a kernel diagram because any feedback morphism b making left triangle commutative must
177 verify im F (b) ⊂ ker a.

(W, g, D) _ _ _ _ _ _ _ _ _ _ _/ (ker a, f, (B ∩ ker a))

∃!
??
?? f = (j) 
?? 
?? 
?? b i 
?
 ??  
?? 

0
??  0
j
??  1

     
f11 f12 b1
coim a ⊕ ker a, f21 f22
, b2

0   0
η 0
a 0 0

)   0 0
u 
b01
 
f11 f12
im a ⊕ coker a, 0 0
f22
, b02
(18)

 
0
178 It is clear that any b making left triangle commutative must be on the form b = j
where j injects W into ker a.
179 Thus it only remains to check that f is a feedback morphism, which is straightforward from b and i are feedback
180 morphisms. 

181 Lemma 11 Let a : σ = (V, f, B) → (V 0 , f 0 , B 0 ) be a feedback morphism, let a = F (a) and suppose that
182 B ≤ ker(a). Let S = {S ≤ ker a : f (S) ⊂ S + B}. Then S is a lattice and Ker(a) = (max S, f, (B ∩ max S))

183 Proof. It is straightforward to prove that S is a lattice of subspaces of ker a; i.e. S =

6 ∅ and if S1 , S2 ∈ S then
184 (S1 ∩ S2 ) ∈ S, and (S1 + S2 ) ∈ S. Since dim ker a < +∞, there exists a maximum of S.
185 Put K = max S and i : K → ker a natural inclusion.The following is a kernel diagram:

(W, g, D) _ _ _ _ _ _ _ _ _ _/ (K, (πK ◦ f ), (B ∩ K))

∃!
??
?? f = (j) 
?? 
?? 
?? b i 
?
 ??  
?? 

0
??  0
j
??  i
 
 
    
f11 f12 b1
coim a ⊕ ker a, f21 f22
, b2

0   0
η 0
a 0 0

)   0 0
u 
b01
 
f11 f12
im a ⊕ coker a, 0 0
f22
, b02
(19)

186 It is clear that any a ◦ i = 0. On the other hand,

 for any b such that a ◦ b = 0, it follows that f making left
0
187 triangle commutative must be on the form f = j
where j injects W into K = max S because j(W ) ∈ S.
188 Thus it only remains to check that f is a feedback morphism, which is straightforward from b and i are feedback
189 morphisms.
190 

8
 The exact structure of feedback actions

191 The Second Reduced Form

192 Consider a FRF, i.e. a feedback morphism a : σ → σ 0 on the form

 
η 0
0 0
/ im a ⊕ coker a,
0 0
b01
         
f11 f12 b1 f11 f12
coim a ⊕ ker a, , im 0 , im b02
f21 f22 b2 a 0 f22
(20)

193 Split vector spaces coim a and Kdim B according to linear map b1 . That is to say, if the Smith reduced form of
194 linear map b1 : Kdim B → coim a is
 
β11 0
0 0
Kdim B = coim b1 ⊕ ker b1 −−−−−−−−→ (im b1 ⊕ coker b1 ) = coim a (21)
( b1 )

195 then we have the following decomposition

 
β11 0
0
 0 
β31 β32
Kdim B = coim b1 ⊕ ker b1 −−−−
−−−− −−→ (im b1 ⊕ coker b1 ) ⊕ ker a (22)
b1
b2

196 and hence linear map f is written

 
h11 h12 h13
 h21 h22 h23 
h31 h32 h33
(im b1 ⊕ coker b1 ) ⊕ ker a −−−
−−−−−−−−−−−→ (im b1 ⊕ coker b1 ) ⊕ ker a (23)
f11 f12
f21 f22

197 Then above feedback morphism a is now, according to the decompositions, on the form

 
! !! η1 η2 0
h11 h12 h13 β11 0 0 0 0
(im b1 ⊕ coker b1 ) ⊕ ker a, h21 h22 h23 , im 0 0 / σ0
h31 h32 h33 β31 β32 a

(24)
198 where β11 : coim b1 → im b1 is the induced isomorphism, and η = (η1!
, η2 ).
1 0 0
199 Now consider the invertible linear map P = 0 1 0 . Put (h0ij ) = P (hij )P −1 . Then, the
−1
−β31 β11 0 1
200 feedback map
! !!
h11 h12 h13 β1 0
(im b1 ⊕ coker b1 ) ⊕ ker a, h21 h22 h23 , im 0 0
h31 h32 h33 β31 β32

 
1 0 0
ϕ  0 1 0 
−1
−β31 β11 0 1

 ! !!
h011 h012 h013 β1 0
(im b1 ⊕ coker b1 ) ⊕ ker a, h021 h022 h023 , im 0 0 (25)
h031 h032 h033 0 β32

9
 The exact structure of feedback actions

201 is an isomorphism of linear systems.

202 Finally, since β11 is invertible, it follows that the identity matrix gives a isomorphism

! !!
h011 h012 h013 β1 0
(im b1 ⊕ coker b1 ) ⊕ ker a, h021 h022 h023 , im 0 0
h031 h032 h033 0 β32

 
1 0 0
ψ  0 1 0 
0 0 1

 ! !!
0 0 0 β1 0
(im b1 ⊕ coker b1 ) ⊕ ker a, h021 h022 h023 , im 0 0 (26)
h031 h032 h033 0 β32

203 Therefore we have just proven the following result

204 Lemma 12 Let a : σ → σ 0 be a feedback morphism. Let a = F (a). Then there exist isomorphisms of linear
205 systems such that

! !!
∼
= 0 0 0 β1 0
(V, f, B) / (im b1 ⊕ coker b1 ) ⊕ ker a, h021 h022 h023 , im 0 0
h031 h032 h033 0 β32

 
η1 η2 0
a 0 0 0

 
∼
= / coim a ⊕ ker a,
    
f11 f12 b1
(V , f 0 , B 0 )
0
f21 f22
, im b2
(27)

206 This form of the feedback morphism a will be called a second reduced form of a or SRF.

207 We need to set a final detail before the main result. Let a be a SRF. Let us define K = max S the maximum of
208 lattice of finite-dimensional vector subspaces of ker h23 ≤ ker a given by

S = {S ≤ ker h23 : h33 (S) ⊆ S + im β32 } (28)

209 that is to say, the greater subspace of ker h23 such that it is h33 -invariant modulo im β32 . Note that one can compute
210 K as K = ker (πim β32 ◦ (h33 − 1)) ∩ ker h23 , but it will be interesting to remark that K is also the maximum of
211 the above lattice.

212 Theorem 13 SK has all kernels

213 Proof. It is sufficient to compute the kernel of a second reduced form (SRF)

10
 The exact structure of feedback actions

! !!
0 0 0 β11 0
im b1 ⊕ coker b1 ⊕ ker a, h21 h22 h23 , im 0 0
h31 h32 f33 0 β32

 
η1 η2 0
a 0 0 0


0 0
b01
    
f11 f12
im a ⊕ coker a, 0
f21 0
f22
, im b02
(29)

214 The following is a kernel diagram

(W, g, D) _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _/ (K, (πK ◦ h33 ) , B ∩ K))

∃!
LLL
LLL (y) rr
LLL rrrr
   
0 0
LLL  0   0 
rrr
LLL y i
r rrrr
LLL r
LLL
rrrrr i
LLL r
% yrrr ! !!
0 0 0 β11 0
im b1 ⊕ coker b1 ⊕ ker a, h21 h22 h23 , im 0 0
h31 h32 h33 0 β32

 
η1 η2 0
a 0 0 0

0 0

0 0
b01
    
f11 f12
, im a ⊕ coker a, 0
f21 0
f22
, im b02 r (30)

215 First, (K, (πK ◦ h33 ) , B ∩ K)) is a linear system because (B ∩ K) ≤ K and πK ◦ h33 : K → K is a linear
216 map.
217 Second,
! ! !!
0 0 0 0 β11 0
i= 0 : (K, (πK ◦ h33 ) , B ∩ K)) → im b1 ⊕ coker b1 ⊕ ker a, h21 h22 h23 , im 0 0
i h31 h32 h33 0 β32
(31)
218 is a feedback morphism because K ≤ ker a and (im β32 ∩ ker h23 ) ≤ im β32 , thence, for all x ∈ K
! ! ! ! ! !
0 0 0 0 0 0 β11 0
h21 h22 h23 0 − 0 (πK ◦ h33 ) (x) = h23 ix ⊆ im 0 0 (32)
h31 h32 h33 i i h33 ix − πk h33 x 0 β32

219 because x ∈ ker h32 and h33 x ∈ K + im β32

220 Finally, universal property
! follows by taking into account that the structure of any morphism annihilating a must
0
221 be on the form 0 because (η1 , η2 ) is an isomorphism. Then y(W ) must lie in a h33 -invariant modulo β32
y
!
0
222 subspace of ker h23 in order to 0 being a feedback morphism, a fortiori y(W ) must lie in K due to maximality
y
223 of K. Analogous argument works to prove that y(D) ⊆ (K ∩ B). Therefore linear map y : W → K defines a
224 feedback morphism y : (W, g, D) → (K, (πK ◦ h33 ) , B ∩ K)) and this morphism is the unique morphism making
225 commutative the whole kernel diagram.
226 

11
 The exact structure of feedback actions

227 Note 14 All equalisers in category SK are obtained as difference kernels, and all coequalisers are obtained as
228 difference cokernels. Hence category SK is both finite complete (there exist all finite limits), and finite cocomplete
229 (there exist all finite colimits).

230 Now we compute the kernel of a feedback morphism of reachable systems in the category of reachable systems
231 AK .

232 Theorem 15 Let a : σ → σ 0 be a morphism of reachable linear systems (in AK ) and suppose that a is SRF. Then
233 its kernel in AK is given by
∗

(πK ◦ h33 ) (B ∩ K), πk ◦ f33 , B ∩ K

 
0
i  0 
i

 ! !!
0 0 0 β11 0
im b1 ⊕ coker b1 ⊕ ker a, h21 h22 h23 , im 0 0
h31 h32 h33 0 β32

 
η1 η2 0
a 0 0 0


0 0
b01
    
f11 f12
im a ⊕ coker a, 0
f21 0
f22
, im b02
(33)
∗
234 that (πK ◦ h33 ) (B ∩ K) ≤ K. On the other hand, any reachable linear
Proof. It suffices to take into account !
0
235 system (W, g, D) and morphism 0 satisfies y(D) ⊂ (B ∩ K). Since (W, g, D) is reachable it follows that
y
∗
236 y(W ) ⊂ (πK ◦ h33 ) (B ∩ K).
237 

238 Note 16 All equalisers in category AK are obtained as difference kernels, and all coequalisers are obtained as
239 difference cokernels. Hence category AK is both finite complete (there exist all finite limits), and finite cocomplete
240 (there exist all finite colimits).

241 The cokernel of a feedback morphism of reachable systems a can be computed in the category of reachable
242 systems or in the category of linear systems with the same result, that is to say, CokerAK a = CokerSK a, this is
243 because the forget-reachability functor I : AK → SK is right-exact. When computing kernels one needs to care
244 about the ambient category of systems.

245 Example 17 Consider the feedback morphism

   D E ( 1 0 )
0 0 1
Σ2 = (K2 , f, B) = K2 , 1 0
, 0
−−−−−−→ (K, 0, K) = Σ1 (34)
a
246 Since f (ker a) = 0 it follows by Lemma 10 that the kernel in SK is given by
h      D Ei
0 0 0 1
KerSK a = (K, 0, 0), 1 : (K, 0, 0) → K2 , 1 0 , 0
(35)

247 while the kernel in AK is given by

h    D Ei
0 0 1
KerAK a = (0, 0, 0), 0 : (0, 0, 0) → K2 , 1 0
, 0
(36)

12
 The exact structure of feedback actions

248 Above example shows that, in general, kerAK 6= kerSK , hence functor I : AK → SK is not left exact (does
249 not preserve kernels). On the other hand, because of universal property of kernels in SK it follows that for every
250 feedback morphism a between reachable linear systems, the natural inclusion j : (πK ◦ h33 )∗ (B ∩ K) → K gives
251 rise a feedback monomorphism j : kerAK a → kerSK a.
252 Kernels are monic, cokernels are epic. The converse is not true in SK : consider the feedback morphism
(1)
253 (K, 0, 0) −−−→ (K, 0, K) = Σ1 (see [18, 4.4]). It is both monic and epic because T (a) = (1) is an isomorphism,
a
254 but a is neither a kernel nor a cokernel in SK . This proves that SK is neither normal nor conormal category.
255 Example 34 shows that AK is not conormal. The morphism
     ( 1 0 )
0 0 1
Σ[2] K2 , 1 0 , 0 −−−−−−→ (K, 0, K) = Σ[1] (37)
a
256 is a epimorphism because T (a) = (1, 0) is onto, but it is not a cokernel in AK . To see this, suppose by contradiction
 
257 that CokerAK (b : σ → Σ[2] ) = Σ[1] , a : Σ[2] → Σ[1] . Then by [24, p. 193]
h i
Coker(Ker(Coker(b))) = Coker(b) = Σ[1] , a : Σ[2] → Σ[1] (38)

258 but, on the other hand, by direct calculation,

  h i
Coker(Ker(Coker(b))) = Coker 0 : (0, 0, 0) → Σ[2] = Σ[2] , 1 : Σ[2] → Σ[2] (39)

259 this is a contradiction, hence a is not a cokernel and AK is not conormal.

260 To conclude we would like to remark that we conjecture that all monics in AK are kernels, i.e. AK is a normal
261 category.

262 Images and coimages of feedback morphisms, pre-abelian property

263 A pre-abelian category is an additive category where exist all kernels and all cokernels. In a pre-abelian category
264 categorical images and coimages of morphisms are defined: let a : σ → σ 0 be a feedback morphism and consider its
265 cokernel map (c : σ 0 → Coker(a)). Then the image of a is given by the kernel of its cokernel map Im(a) = Ker(c).
266 The coimage is given by the cokernel of its kernel, i.e. if (i : Ker(a) → σ) is the kernel map then coimage is
267 Coim(a) = Coker(i).
268 Feedback morphism a : σ → σ 0 induces a feedback morphism ā : Coim(a) → Im(a). If this latter morphism
269 were always an isomorphism then every feedback morphism would have a canonical decomposition and the cate-
270 gory would be abelian. But this is not the case. A weaker version is that the induced morphism ā were a bimorphism
271 (mono and epic) for all feedback morphism a, and the category would be quasi-abelian. But this is not the case for
272 categories SK and AK . We put an example of a feedback morphism a such that ā is not even bimorphism.
! !!
0 0 0 1
273 Example 18 Consider the reachable linear systems Σ3 = K3 , 1 0 0 , 0 and Σ1 = (K, 0, K).
0 1 0 0
274 The cokernel system of feedback morphism
! !!
0 0 0 1 ( 0 1 0 )
3 3
Σ = K , 1 0 0 , 0 −−−−−−−−−→ (K, 0, K) = Σ1 (40)
0 1 0 0 ā

275 is Coker(a) = (0, 0, 0), while the cokernel map is c = 0 : (K, 0, K) → (0, 0, 0). Hence image of morphism a is
 
276 given by Im(a) = Ker (c = 0 : (K, 0, K) → (0, 0, 0)) = (K, 0, K), 1(K,0,K) .
277 On the other  by Lemma 11, the kernel system is (K, 0, 0), and the kernel morphism is given by natural
 hand,
0
278 inclusion i= 0 : (K, 0, 0) → Σ3 . Hence the coimage of feedback morphism a is Coim(a) = Coker(i) =
1


279 K2 , 0, K . Finally, induced feedback morphism ā is the following

( 0 1 )
Coim(a) = K2 , 0, K −−−−−−→ (K, 0, K) = Im(a) (41)
ā

13
 The exact structure of feedback actions

280 which is not a monomorphism, just by taking into account that ā ◦ bλ = 0 for all element in the family of morphisms
      
bλ = λ0 : (K, 0, 0) → K2 , 01 00 , 10 (42)

281 A Exact structure in the category of linear systems

282 Below we recall the definition of exact category in the sense of Quillen (cite).

Definition 19 (cf. [20] or [21]) A kernel-cokernel pair in SK (respectively AK ) is a pair of feedback morphisms
a b
σ 0 −−−→ σ −−−→ σ 00

283 where a is the kernel of b and b is the cokernel of a.

284 If a class E of kernel-cokernel pairs is fixed, an admissible monic is a morphism i such that there exist a
285 morphism p such that (i, p) ∈ E. An admissible epic is defined dually. We depict admissible monics by ,−−→ and
286 admissible epics by −−−→→

287 Definition 20 (cf. [20] or [21]) An exact structure is a class E of kernel-cokernel pairs which is closed under
288 isomorphisms of linear systems and satisfies the following axioms:

289 i The identity of the zero object 1(0,0,0) is an admissible epic.

290 ii The class of admissible monics is closed under composition. The class of admissible epics is closed under
291 composition.
i a
292 iii The push-out of an admissible monic σ ,−−→ τ along an arbitrary feedback map σ −−→ σ 0 exists and yields an
293 admissible monic.

σ
i /τ



a 

 
0 _ _ _/ σ 0 t τ (43)
σ σ

p a
294 iv The pull-back of an admissible epic σ −−−→→ τ along an arbitrary feedback map τ 0 −−→ τ exists and yields an
295 admissible epic.
σ ×τ τ 0 _ _ _/ / τ 0

 a


 p 
σ //τ (44)

296 There are several exact structures on an additive category. However, the kernel-cokernel pairs isomorphic to
 
1
0 ( 0 1 )
σ ,−−−−→ σ ⊕ τ −−−−−−−
→→τ (45)

297 form a exact structure Emin , and every other exact structure contains Emin (see [20, Proposition 2.12]).

298 Theorem 21 The minimal exact structure on categories of linear systems SK or AK contains all kernel-cokernel
299 pairs on the form section-retraction, i.e. all kernel-cokernel pairs on the form
a b
σ 0 ,−−−→ σ −−−→→ σ 00 (46)

14
 The exact structure of feedback actions

300 Proof.
301 

302 Theorem 22 (Kalman’s Decomposition) Let σ = (V, f, B) be a linear system in SK . Define σr = (f ∗ (B), f, B)
303 and σ0 = (V /f ∗ (B), f¯, 0) then σ ∼
= σr ⊕ σ0 .

304 Proof. Since f ∗ (B) = B + f (B) + f 2 (B) + · · · is f invariant and B ≤ f ∗ (B) it follows that both σr and
305 σ0 are in SK . The natural inclusion i : f ∗ (B) → V gives rise a feedback monomorphism i : σr → σ while the
306 natural quotient map p : V → V /f ∗ (B) gives rise a feedback epimorphism p : σ → σ0 . Moreover, the following
307 is a kernel-cokernel pair in SK
i p
σr0 −−−→ σ −−−→ σ0 (47)
308 In fact, above kernel-cokernel pair is an exact sequence in Emin . By [21, Lemma 7.1], it suffices to prove that either
309 i is a section or p is a retract. But both conditions are clear just by splitting V ∼
= f ∗ (B) ⊕ V /f ∗ (B). Therefore the
310 result.
311 

312 Theorem 23 (Brunovsky’s Theorem) Let σ = (V, f, B) a reachable linear system (in AK ) and set n = dim V .
313 Then there exist a partition κ1 + · · · + κp = n of integer n such that σ = Σκ1 ⊕ · · · ⊕ Σκp

314 Proof.- By recursion in n. Consider a non zero vector ~v1 ∈ B and linear system σ1 = (f ∗ (< ~v1 >), f, < ~v1 >).
315 Natural inclusion ι1 : f ∗ (< ~v1 >) → V gives rise a exact sequence in AK

i p
σ1 = (f ∗ (< ~v1 >), f, < ~v1 >) −−−→ σ −−−→ V /f ∗ (< ~v1 >), f¯, F (p)(B)


(48)

Now σ ∼= σ1 ⊕ V /f ∗ (< ~v1 >), f¯, F (p)(B) . System V /f ∗ (< ~v1 >), f¯, F (p)(B) is reachable, and note that



316

317 dim (V /f ∗ (< ~v1 >)) < n. Thence, by recursion, σ ∼ = σ1 ⊕ · · · ⊕ σp where all σi are in AK .
318 To conclude the proof note that every system σi = (Vi , f, < vi >) is isomorphic to a system on the form Σκi
319 because of dim < vi >= 1.
320 

321 Simple reachable systems

322 Consider the exact category (AK , Emin ) of reachable linear systems and short exact sequences (section-retraction
323 pairs). We say that σ 0 is an admissible subsystem of σ (denoted by σ 0 ⊂ σ) if there exists an admissible section
i i
324 σ 0 ,−−−→ σ [20, 3.1.], or equivalently if one has the short exact sequence of reachable systems σ 0 ,−−−→ σ −−−
→ σ 00
→
325 for some σ 00 .

326 Definition 24 ([20]) A nonzero system σ is simple if every subsystem σ 0 is either zero or isomorphic to σ. Simple
327 objects are indecomposable.

328 Theorem 25 Consider the category of reachable systems and feedback morphisms equipped with the minimal
329 exact structure Emin of section-retraction pairs. Then a system σ is Emin -simple if and only if it is isomorphic to a
330 Canonical Controller Form, i.e. σ ∼ = Σn .
i r
331 Proof. Suppose, that σ 0 ,−−−→ Σn −−− → σ 00 is short exact. Consider the one-side inverses q ◦ i = 1σ0 and
→
332 r ◦ j = 1σ00 . Then, by [21, Remark 7.4.], feedback morphism ( i j ) : σ0 ⊕ σ00 → Σn is a isomorphism with
inverse
q .
333
r
0 00
associated to Σn satisfies ZkΣ ∼
n
n
334 Then the sequence of invariant vector spaces ZkΣ k≥1
= Zkσ ⊕ Zkσ (see [15,
n 0 00
335 Lemma 3.6.]). In particular, dim ZkΣ = dim Zkσ + dim Zkσ in N. But since the sequence of invariants of Σn is

15
 The exact structure of feedback actions

in fact ZkΣ = 0 if k 6= n and ZnΣ = K, it follows that either σ 0 ∼

= Σn and σ 00 = (0, 0, 0), or σ 0 ∼
n n
336 = (0, 0, 0) and
00 n
337 σ =Σ
338 

339 Note 26 Recall that EndAK (Σn ) ∼ = K (see [16, 6.3.]) hence simple systems on the form Σn are bricks [19,
340 Definition 2.1]. Therefore every simple object in AK is a brick, and first statement of Schur’s Lemma [19, Lemma
341 2.3.] is fulfilled. Second statement of Schur’s Lemma does not hold in AK because homAK (Σ2 , Σ1 ) 6= 0 though Σ2
342 and Σ1 are both simple.

343 Note 27 To conclude we remark that Brunovsky’s Theorem implies that every reachable system is decomposed in
344 a biproduct of Emin -simple objects. Hence category AK is object semisimple.

345 Conclusions
346 The categories of linear systems and of reachable linear systems over finite vectorial spaces are pre-abelian cate-
347 gories. Kernels and cokernels (an thus images and coimages) of feedback morphisms can be computed effectively.
348 The minimal exact structure is introduced: the short exact sequences are the section-retract pairs. This minimal
349 exact structure is sufficient to state and understand classical decomposition theorems of linear systems (Kalman’s
350 Decomposition, Brunovsky’s Theorem) in terms of short exact sequences of linear systems.
351 Future works would include a traslation of our results to category of regular systems over a commutative ring,
352 and the study of maximal exact structure in categories of linear systems.

353 Acknowledgments
354 The author would like to thank CAFE (Ciberseguridad, Aplicaciones, Fundamentos y Educación) research group
355 for its support.

356 Conflict of interest

357 The author declares no potential conflict of interests.

358 References
359 [1] J.C. Maxwell. “On governors”. vol. .

360 [2] K. Ogata. Modern control engineering. Prentice Hall, 2010.

361 [3] J.L. Casti. Linear dynamical systems. Academic, 1987.

362 [4] J.W. Brewer, J.W. Bunce and F.S VanVleck. Linear dynamical systems over commutative rings. Marcel
363 Dekker, 1986.

364 [5] E.D. Sontag. Mathematical Control Theory. Springer, 1991.

365 [6] N. DeCastro-García. “Feedback equivalence of convolutional codes over finite rings”. Open Math., vol. 15,
366 2017.

367 [7] J. Ferrer, M.I. García-Planas and F. Puerta. “Brunovsky local form of a holomorphic family of pairs of
368 matrices”. Lin. Algebra App., vol. 253, 175–198, 1997.

16
 The exact structure of feedback actions

369 [8] J.A. Hermida-Alonso, M.P. Pére and T. Sánchor-Giralda. “Brunovsky canonical form for linear dynamical
370 systems over commutative rings”. Lin. Algebra App., vol. 233, 131–147, 1996.

371 [9] A.L. Muñoz Castañeda, N. DeCastro-García and M.V. Carriegos. “On the state approach representations of
372 convolutional codes over rings of modular integers”. Mathematics, vol. 9, 2962, 2021.

373 [10] J.M. Muñoz Porras and J.I. Iglesias Curto. “Classification of convolutional codes”. Lin. Algebra App., vol.
374 432, 2701–2725, 2010.

375 [11] V. Herranz, D. Napp and C. Perea. “Serial concatenation of a block code and a 2d convolutional code”.
376 Multidimensional Systems and Signal Processing, vol. 30, no. 3, 1113–1127, 2019.

377 [12] D. Napp, R. Pinto and C. Rocha. “State representations of convolutional codes over a finite ring”. Lin. Algebra
378 App., vol. 640, 48–66, 2022.

379 [13] R.E. Kalman. Kronecker invariants and feedback, In: Ordinary differential equations. Academic, 1972.

380 [14] P.A. Brunovsky. “A classification of linear controllable systems”. Kibernetika, vol. 3, 1970.

381 [15] M.V. Carriegos and A.L. Muñoz Castañeda. “On the K-theory of feedback actions on linear systems”. Lin.
382 Algebra App., vol. 440, 233–242, 2014.

383 [16] M.V. Carriegos. “Morphisms of linear control systems”. Open Math., vol. Submitted, 2022.

384 [17] M.V. Carriegos. “Enumeration of classes of linear systems via equations and via partitions in a ordered abelian
385 monoid”. Lin. Algebra App., vol. 438, 2013.

386 [18] M.V. Carriegos. “Some categorical properties of linear systems”. Mathematics, vol. Submitted, 2022.

387 [19] H. Enomoto. “Schur’s lemma for exact categories implies abelian”. J. Algebra, vol. 584, 260–269, 2021.

388 [20] T. Brüstle, S. Hassoun, D. Langford and S. Roy. “Reduction of exact structures”. arXiv, vol. math.RT,
389 1809.01282v4, 2021.

390 [21] T. Bühler. “Exact categories”. Expo. Math, vol. 28, no. 1, 1–69, 2010.

391 [22] J.L. Massey and T. Mittelholzer. “Codes, automata, and continuous systems: explicit interconnections”. IEEE
392 Trans. Auto. Control, vol. 12, 644–650, 1967.

393 [23] J.W. Brewer and L. Klingler. “On feedback invariants for linear dynamical systems”. Lin. Algebra App., vol.
394 325, 2001.

395 [24] S. Mac Lane. Categories for the working mathematician. Springer, 1971.

17