the acyclic matrices with a p set of maximum size 2015 linear algebra and its applications

Upload: takeshi-kazuo

Post on 06-Jul-2018

213 views

Category:

Documents


0 download

TRANSCRIPT

  • 8/17/2019 The Acyclic Matrices With a P Set of Maximum Size 2015 Linear Algebra and Its Applications

    1/11

    Linear Algebra and its Applications 468 (2015) 27–37

    Contents lists available at  ScienceDirect

    Linear Algebra and its Applications

    www.elsevier.com/locate/laa

    The acyclic matrices with a P-set of maximumsize

    Zhibin Du a, Carlos M. da Fonseca b,∗

    a School of Mathematics and Information Sciences, Zhaoqing University, Zhaoqing 526061, Chinab Department of Mathematics, Kuwait University, Safat 13060, Kuwait 

    a r t i c l e i n f o a b s t r a c t

     Article history:

    Received 16 June 2013

    Accepted 22 October 2013

    Available online 23 December 2013

    Submitted by P. Psarrakos

    MSC:15A18

    15A48

    05C50

    Keywords:

    Trees

    Acyclic matrices

    Multiplicity of eigenvalues

    P-set

    P-vertices

    Let   m A (0)   denote the nullity of a given matrix   A   of order   n. Set

     A(α)   for the principal submatrix of   A  obtained after deleting therows and columns indexed by the nonempty subset  α  of  {1, . . . ,n}.When  m A(α)(0) = m A (0) + |α|, we call  α  a P-set of   A. In this paper,we classify all of the trees   T   for which there exists a matrix   A

    whose graph is   T   and containing a P-set of maximum size. Our

    characterization does not depend on whether the acyclic matrices

    are singular or nonsingular.

    ©  2013 Elsevier Inc. All rights reserved.

    1. Introduction

    Let   A =  (ai j )   be an   n × n   real symmetric matrix. The graph of   A, denoted by   G( A), is defined tobe the graph whose vertex set is   {1, . . . ,n}   and edge set is   {i j | i      =   j   and  ai j      = 0}. In this sense,   G( A)does not depend on the main diagonal entries of   A. Here we focus our study on the set

    S (G) =

     A ∈ Rn×n   A is symmetric and  G ( A) = G,

    where   G   is a given tree on  n  vertices. Any matrix   A ∈ S (G)  is called  acyclic .

    *   Corresponding author.E-mail addresses: [email protected] (Z. Du), [email protected] (C.M. da Fonseca).

    0024-3795/$ – see front matter   © 2013 Elsevier Inc. All rights reserved.

    http://dx.doi.org/10.1016/j.laa.2013.10.042

    http://dx.doi.org/10.1016/j.laa.2013.10.042http://www.sciencedirect.com/http://www.elsevier.com/locate/laamailto:[email protected]:[email protected]://dx.doi.org/10.1016/j.laa.2013.10.042http://crossmark.crossref.org/dialog/?doi=10.1016/j.laa.2013.10.042&domain=pdfhttp://dx.doi.org/10.1016/j.laa.2013.10.042mailto:[email protected]:[email protected]://www.elsevier.com/locate/laahttp://www.sciencedirect.com/http://dx.doi.org/10.1016/j.laa.2013.10.042

  • 8/17/2019 The Acyclic Matrices With a P Set of Maximum Size 2015 Linear Algebra and Its Applications

    2/11

    28   Z. Du, C.M. da Fonseca / Linear Algebra and its Applications 468 (2015) 27–37 

    The principal submatrix of   A   obtained after deleting the rows and columns indexed by   α  ⊂{1, . . . , n}   is denoted by   A(α). If we consider the principal submatrix of   A   consisting of both rowsand columns indexed by  α, then we will write   A[α]. However, we reserve the notation   A(G)   for thestandard adjacency   (0, 1)-matrix of the graph  G .

    Given a tree   T   with vertices 1, . . . , n, for any nonempty subset   α   of   {1, . . . , n}, we denote byT (α)  the forest obtained from   T   by deleting the vertices in  α   and the incident edges, and   T [α]   thesubgraph (forest) of   T   induced by the vertices in   α. Clearly, if   A  ∈ S (T ), then   A(α) ∈ S (T (α))   and A[α] ∈ S (T [α]), for any nonempty subset  α   of  {1, . . . , n}.

    Denoting by  m A (λ)   the (algebraic) multiplicity of an eigenvalue  λ  of   A, we assume that  m A (λ) = 0if   λ   is not an eigenvalue of   A. As a consequence of the Cauchy Interlacing Theorem   [9, Theorem

    4.3.17], we have

    m A (λ) − 1 m A(i)(λ)m A (λ) + 1.   (1.1)

    In the case of equality in the right-hand side of  (1.1), with   λ = 0, the index   i   is known as a  P-vertexof   A  [10,11].  Notice that  m A (0)  corresponds to the nullity of   A.

    One may also derive from the right-hand side of  (1.1) that

    m A(α)(0) m A (0) + |α|,   (1.2)

    for any nonempty subset   α   of   {1, . . . , n}. When we have equality in   (1.2), we call   α   a   P-set   of   A[10].   Therefore,   α   is a P-set of   A   if and only if the nullity of   A(α)   is   |α|  more than that of   A. Themaximum size of a P-set of   A   is denoted by   P s( A)   [12].   Interestingly, from the Cauchy Interlacing

    Theorem, every subset of a P-set of   A   is also a P-set of   A. In particular, each vertex in a P-set of   A   is

    a P-vertex of   A   [12, Proposition 5]  (see also  [10] for proofs).

    It is also worth mentioning that, when  α   is a P-set of   A, for each proper subset  β ⊂ α   (which is aP-set of   A),  α \ β   is a P-set of   A(β) .

    Motivated by   [11],   in 2009, Kim and Shader   [12]   started a thorough study of the P-vertices and

    P-sets of nonsingular matrices whose graph is a path or a star. Several interesting questions wereproposed there and all answered in [1–6] with some significant results in the more recent references.

    Some considerations on P-sets were also made by Kim and Shader regarding slightly more general

    singular acyclic matrices. Namely, they constructed the nonsingular tridiagonal matrices of order   n

    with P-sets of cardinality   n/2, where    x   stands for the largest integer not greater than   x. Theseauthors verified this (tight) upper bound for nonsingular acyclic matrices and for singular acyclic

    matrices of odd order. But, in fact, this is the maximum size of any P-set of any symmetric matrix of 

    order n . Our result is motivated by  [12, p. 402] which is apparently proved for all symmetric matrices.

    It is stated by its importance in this paper and the proof is included for completeness.

    Proposition 1.1. For any symmetric matrix A of order n, we have

    P s( A)

    n

    2

    .

    Proof.   Suppose that  α  is a P-set of   A, i.e.,  m A(α)(0) = m A (0) + |α|. On one hand,

    m A(α)(0) = m A (0) + |α| |α|.   (1.3)

    On the other hand, noting that the matrix   A(α)   is of order  n − |α|, we have

    m A(α)(0) n − |α|.   (1.4)

    Now it follows that   |α| n − |α|, i.e.,

    |α|

    n

    2

    .

    The result is now clear.  

  • 8/17/2019 The Acyclic Matrices With a P Set of Maximum Size 2015 Linear Algebra and Its Applications

    3/11

     Z. Du, C.M. da Fonseca / Linear Algebra and its Applications 468 (2015) 27–37    29

    Finally, it is worth mentioning that for any tree   T , there exists a nonsingular matrix   A  ∈  S (T )without P-vertex, so   P s( A) = 0. Moreover, the only tree  T  for which each nonsingular matrix   A ∈ S (T )satisfies   P s( A) 1 is the star [1].

    The thrust of this paper is to completely characterize the trees where the bound of  Proposition 1.1

    is attained. Our classification only depends on the parity of the order of the matrices. Before ourmain results, we will establish several technical lemmas which may be of independent interest. We

    will also recall some previous results.

    2. Preliminary results

    Let us denote by   P n   the path on  n  vertices.

    Our first result observes that the only tree of order  n   for which there exists a singular matrix of 

    nullity n − 1 is the path   P 2 .

    Lemma 2.1. Suppose that T is a tree on n 2 vertices, and A ∈ S (T ). If m A (0) = n − 1, then T  ∼= P 2 .

    Proof.   Observe that  m A (0) = n − 1 1 implies that the spectrum of   A  consists of two distinct eigen-values, one is zero, and the other is nonzero. Thus 0 is either the largest or the smallest eigenvalue

    of   A.

    It is well known that whether the largest or the smallest eigenvalue of an acyclic matrix is simple

    (see, e.g.,   [10, Corollary 7]). So  m A (0) = n − 1 = 1, i.e.,  n = 2. Accordingly,   T  ∼= P 2 .  

    The next theorem asserts an important property for the nullity of an acyclic matrix with a P-set

    of maximum size: it is at most 1. We recall that a trivial graph contains only one vertex and, conse-

    quently, no edges. If, in addition, its nullity is 1, then this will mean that the weight of such vertex

    is 0.

     Theorem 2.2. Let T be a tree on n 2 vertices, and A ∈ S (T ). If  α  is a P-set of A with  |α| = n2

    , then one of the following conditions must be satisfied:

    (i)  n is even, m A (0) = 0, and each component of T (α) is trivial and of nullity 1.(ii)  n is odd, m A (0) = 1, and each component of T (α) is trivial and of nullity 1.

    (iii)  n is odd, m A (0) = 0, and each component of T (α) is trivial and, with one exception, of nullity 1.(iv)  n is odd, m A (0) = 0, T (α) contains exactly one edge, and each component of T (α) is of nullity 1.

    Proof.   Let  T̄  = T (α)  and   ¯ A =  A (α) ∈ S (T̄ ).

    Case 1.  n  is even, i.e.,   |α| =  n2  and, consequently,   ¯ A   is of order

      n2 .

    From the inequalities (1.3) and (1.4), we have

    n

    2 m A (0) +

     n

    2 = m ¯ A (0) n −

     n

    2 =

     n

    2.

    Thus  m A (0) = 0 and  m ¯ A (0) =  n2 . Since

      ¯ A   is symmetric, the rank of   ¯ A   is 0, i.e.,   ¯ A   is the zero matrix.

    Accordingly,  T̄   is an edgeless graph. So (i) follows.

    Case 2.  n   is odd, i.e.,   |α| =   n−12

      and, consequently,   ¯ A   is of order   n+12

      .

    From the inequalities (1.3) and (1.4), we have

    n − 1

    2  m A (0) +

     n − 1

    2  = m ¯ A (0) n −

     n − 1

    2  =

     n + 1

    2  .

    Thus m A (0) = 0 or 1.

  • 8/17/2019 The Acyclic Matrices With a P Set of Maximum Size 2015 Linear Algebra and Its Applications

    4/11

  • 8/17/2019 The Acyclic Matrices With a P Set of Maximum Size 2015 Linear Algebra and Its Applications

    5/11

     Z. Du, C.M. da Fonseca / Linear Algebra and its Applications 468 (2015) 27–37    31

    Proof.   Let us set  β = {u, v, w} ∩ α   and   = |β|. Suppose that (i) does not hold, i.e.,       = 1.If    = 0, then  u, v, w  are in the same component of   T (α), which is a contradiction to the hypoth-

    esis that each component of   T (α)  contains at most one edge. Thus       = 0.If  {u, v} ⊆ α, then {u, v} would be a P-set of   A, which implies that  u  is a P-vertex of   A(v), however,

    which is impossible whether   A[u] = 0 or not, since   u   is an isolated vertex of   T (v). This implies that     = 3, and if   = 2, then   β = {u, w}  or  β = { v, w}.

    Let  ᾱ = α ∩ V (T̄ ), i.e.,  ᾱ = α \ {u, v}, and let   ¯ A =  A[T̄ ].Clearly, |ᾱ| = |α| − 1, whether  β = {u, w}  or  β = { v, w}. Now, since  ᾱ ⊆ α, we have  ᾱ   is a P-set of 

     A, i.e.,  m A(ᾱ)(0) = m A (0) + |ᾱ|.Suppose that   β  = {u, w}. Then   u   is a P-vertex of   A(w), equivalently,   u   is a P-vertex of   A[u, v].

    Thus   A[v] =  0 and   m A[u,v](0) =  0. On one hand, since   u   is a P-vertex of   A, from   Lemma 2.4(i), wehave  m A (0) = m ¯ A (0). On the other hand, taking into account that   w ∈  ᾱ, we have

    m A(ᾱ)(0) = m A[u,v](0) + m ¯ A(ᾱ)(0) = m ¯ A(ᾱ)(0).

    So we get that

    m ¯ A(ᾱ)(0) = m A(ᾱ)(0) = m A (0) + |ᾱ| = m ¯ A (0) + |ᾱ|,

    i.e.,  ᾱ   is a P-set of   ¯ A.If   β = {v, w}, then we can deduce that  ᾱ  is a P-set of   ¯ A   similarly.In particular, if   T   has  n  vertices, and   P s( A) =

    n2 , then, from Proposition 1.1,   P s(

    ¯ A) = n−22   .  

    Lemma 3.2. Suppose that T is a tree with a pendant P 3 =  uvw , and A ∈ S (T ). Let  T̄ be the tree obtained from T by deleting the vertices u, v. Moreover, let us assume that  α  is a P-set of A such that each component of T (α) is trivial and of nullity  1. Then |α ∩ V (T̄ )| = |α| − 1, and α ∩ V (T̄ ) is a P-set of A[T̄ ]. In particular, if T has n vertices, and P s( A) =

    n2 , then P s( A[T̄ ]) =

    n−22   .

    Proof.   Let   β   and    be defined as previously in the proof of   Lemma 3.1.

    If       = 1, then from  Lemma 3.1, the result follows.Clearly, the hypothesis that each component of   T (α)   is trivial implies that if   = 1, then  β = {v}.Suppose in the following that   β  = { v}. Then the hypothesis that each component of   T (α)   is of 

    nullity 1 implies that   A[u] = 0.Let us set  ᾱ = α ∩ V (T̄ ), i.e.,  ᾱ = α \ {u, v}, and let   ¯ A =  A [T̄ ]. Clearly,   |ᾱ| = |α| − 1.Since  ᾱ  ⊆  α , we have  ᾱ   is a P-set of   A, i.e.,   m A(ᾱ)(0) =  m A (0) + |ᾱ|. Recall that   A[u] = 0 and,

    from   Lemma 2.4(ii), we can get   m A (0) = m ¯ A (0)   and   m A(ᾱ)(0) = m ¯ A(ᾱ)(0). Now we have   m ¯ A(ᾱ)(0) =

    m ¯ A (0) + |ᾱ|, i.e.,  ᾱ  is a P-set of   ¯ A.

    In particular, if   T   has  n  vertices, and   P s( A) = n

    2

    , then, from Proposition 1.1,   P s( ¯ A) = n−2

    2

      .  

    We now recall a key definition in  [4]:

    Definition 3.2.   If  x, y   are two terminal vertices sharing a common vertex   z   in a tree   T , then the path

    of   T   induced by the vertices  x, y, z   is said to be  an outer P 3   of T . In this case, we write   P 3 = xz y.

    Observe that we are identifying   P 3 = xzy   with the star   S 3   on 3 vertices where   z   is the centralvertex.

    Lemma 3.3. Suppose that T is a tree with an outer P 3 = xzy, and A ∈ S (T ). If  α is a P-set of A such that each

    component of T (α) contains at most one edge, then exactly one of x, y, z is in  α .

    Proof.   Let us set  γ  = { x, y, z } ∩ α   and   = |γ |.If   = 0, then  x, y, z  are in the same component of  T (α), which is a contradiction to the hypothesis

    that each component of   T (α)  contains at most one edge. Thus       = 0.

  • 8/17/2019 The Acyclic Matrices With a P Set of Maximum Size 2015 Linear Algebra and Its Applications

    6/11

    32   Z. Du, C.M. da Fonseca / Linear Algebra and its Applications 468 (2015) 27–37 

    Observe that   x   is an isolated vertex of   T ( z ), thus whether   A[ x] =  0 or not,   x   would not be aP-vertex of   A( z ), hence  { x, z }  would not be a P-set of   A. Similarly,  { y, z }  would not be a P-set of   A.

    So it follows that       = 3, and if   = 2, then  γ  = { x, y}.Suppose that   γ   = { x, y}. Then   { x, y}   would be a P-set of   A, i.e.,   m A( x, y)(0)  = m A (0)  +   2. So

    m A( x, y, z )(0) m A( x, y)(0) − 1 = m A (0) + 1. On the other hand, since   x ∈ α,   x   is a P-vertex of   A, fromLemma 2.4(i), we have   m A( x, z )(0) = m A (0). Furthermore, note that   y   is an isolated vertex of   T ( x, z ),thus m A( x, y, z )(0) = m A( x, z )(0)  or m A( x, z )(0) − 1, i.e.,  m A( x, y, z )(0) = m A (0)  or  m A (0) − 1, either of which

    is a contradiction to  m A( x, y, z )(0) m A (0) + 1. Thus        = 2.Combining the above arguments, we have   = 1.  

    Lemma 3.4. Suppose that T is a tree with an outer P 3 =  xzy, and A ∈ S (T ). If  α is a P-set of A such that eachcomponent of T (α) is trivial and of nullity 1, then A is singular.

    Proof.   Defining  γ   and    as previously in the proof of   Lemma 3.3.From  Lemma 3.3, we have    =  1. The assumption that each component of   T (α)   is trivial implies

    that  γ   = { z }.If   γ   = { z }, then the hypothesis that each component of   T (α)   is of nullity 1 implies that   A[ x] =

     A[ y] = 0. So   A   is singular.  

    We will call each component obtained from the deletion of a given vertex  v  of a tree  T   as a  branch

    o f T a t v .

    Lemma 3.5. Suppose that T is a tree with an outer P 3 =  xzy, and A ∈ S (T ). Denote by z 1, . . . , z r  the neigh-bors of z in T different from x, y, and T i , for  1 i r, the branch of T at z containing z i . In addition, let  α  bea P-set of A. If z  /∈ α, m A (0) = 0, and each component of T (α) contains at most one edge and of nullity  1, thenα ∩ V (T i ) is a P-set of A [T i ], for  1 i r.

    Proof.   We only prove the result for   i  =  1, i.e.,   α ∩  V (T 1)   is a P-set of   A[T 1], since other cases aresimilar.

    Let αi  = α ∩ V (T i )  and   A i =  A[T i ], for 1 i r .From Lemma 3.3,   exactly one of  x, y, z   is in  α. Assume that  x ∈ α   and   y   /∈ α, since   z  /∈ α.Note that   x ∈ α   implies that   x   is a P-vertex of   A, from   Lemma 2.4(i),   m A( x, z )(0) = m A (0) = 0. So

    m A1 (0) = 0, because   T 1   is a component of   T ( x, z ).Observe that   yz   is   an edge of   T (α), so the hypothesis that each component of   T (α)  contains at

    most one edge implies that   { z 1, . . . , z r } ⊆ α. Hence  { z 1, . . . , z r }   is a P-set of   A, i.e.,

    m A( z 1,..., z r )

    (0) = m A

    (0) + r .   (3.1)

    Since   { x, z 1, . . . , z r } ⊆ α,   x   is a P-vertex of   A( z 1, . . . , z r ), equivalently,   x   is a P-vertex of   A[ x, y, z ],i.e.,  m A[ y, z ](0) = m A[ x, y, z ](0) + 1. Moreover, the hypothesis that each component of   T (α)   is of nullity1 implies that  m A[ y, z ](0) = 1, and thus   A[ x, y, z ]  is nonsingular.

    Note that   α1  ⊆ α   is a P-set of   A. As a consequence,   α1 \ { z 1}   is a P-set of   A( z 1), equivalently,α1 \ { z 1}   is a P-set of   A1( z 1). If   z 1   is a P-vertex of   A1 , then clearly  α1  would be a P-set of   A1 . Thuswe are left to show that   z 1   is a P-vertex of   A1, i.e.,  m A1( z 1)(0) = m A1 (0) + 1 = 1.

    Clearly,

     A( z 1, . . . , z r ) = A[ x, y, z ] ⊕ A1( z 1) ⊕ · · · ⊕  A r ( z r ),   (3.2)

    where  ⊕   represents the direct sum of matrices.From (3.1) and (3.2), we have

    r k=1

    m Ak ( z k )(0) = m A (0) + r  = r .   (3.3)

  • 8/17/2019 The Acyclic Matrices With a P Set of Maximum Size 2015 Linear Algebra and Its Applications

    7/11

     Z. Du, C.M. da Fonseca / Linear Algebra and its Applications 468 (2015) 27–37    33

    Suppose that there exists an index   k, for 1 k r , such that   m Ak( z k )(0) 2. Note that   z k   is a

    P-vertex of   A, i.e.,   m A( z k )(0) = m A (0) + 1 = 1. On the other hand, observe that   T k( z k)   is the vertex-disjoint union of all the components of   T ( z k), except the one containing   z , and thus,

    1 = m A( z k )(0) m Ak ( z k )(0) 2,which is a contradiction.

    So we have   m Ak ( z k)(0) 1, for 1 k   r . Furthermore, together with  (3.3),   we can state that

    m Ak( z k)(0) = 1, for 1 k r . In particular,  m A1( z 1)(0) = 1.  

    4. The characterization for acyclic matrices of even order

    Attaching the path   P 2  to a vertex of a tree   T   is called  adding a pendant P 2   to T . The operation of 

    adding pendant   P 2   to a tree was crucial in our previous works. Interestingly, it also plays a funda-

    mental role in this paper.

    Lemma 4.1. Let T be a tree. Suppose that  T̂ is a tree obtained from T by adding a pendant P 2 =  u v, wherev is of degree  2. If  α  is a P-set of A(T ), then α  ∪ {v} is a P-set of A(T̂ ). In particular, if T has n vertices, andP s( A(T )) =

    n2

    , then P s( A(T̂ )) = n+2

    2  .

    Proof.   Set   A =  A(T )  and   ˆ A =  A(T̂ ).Since  α   is a P-set of   A, we have  m A(α)(0) = m A (0) + |α|.

    Note that   ˆ A[u] = 0 and, from  Lemma 2.4(ii),  m A (0) = m ˆ A (0).

    Observe also that   T   is a component of  T̂ (v), we have   ˆ A(α ∪ {v}) =  A(α) ⊕ ( 0 ). So we can get theequalities

    m ˆ A(α∪{v})(0) = m A(α)(0) + 1

    = m A (0) + |α| + 1

    = m ˆ A (0) +α ∪ {v},

    i.e.,  α ∪ {v}  is a P-set of   ˆ A.In particular, if   T   has  n  vertices, and   P s( A) =

    n2 , then, from Proposition 1.1,   P s(

    ˆ A) = n+22   .  

    Contrary to the operation of adding pendant   P 2 , we can define the operation of deleting pen-

    dant   P 2 . Given a tree   T , deleting a path   P 2 , with degrees 1 and 2 in   T , is called  deleting a pendant P 2of T .

    It is easily verified that if there is no outer   P 3   in a tree   T   on  n 3 vertices, then some deletingpendant   P 2  operation can be applied to   T .

    We are ready to present our first main characterization.

     Theorem 4.2. Let T be a tree on n 2 vertices, where n is even. The following two conditions are equivalent :

    (a)  There exists a matrix A  ∈ S (T ) such that P s( A) =  n2 .

    (b)  T is a tree obtained from P 2 by sequentially adding pendant P 2’s.

    Proof.   (a)  ⇒  (b) Suppose that   A   is a matrix in  S (T )   such that   P s( A) =  n2 .

    Let us start from   T , deleting pendant   P 2’s repeatedly, until there exists an outer   P 3   or a path   P 2

    left, the resulting graph is denoted by   T ∗.If   T ∗ ∼=  P 2, then (b) follows clearly.Suppose that there exists an outer   P 3   in   T 

    ∗ . By   Theorem 2.2(i) and   Lemma 3.2,   we reach that

    m A[T ∗](0) = 0, and there is a P-set  α∗ of   A[T ∗]  such that each component of   T ∗(α∗)   is trivial and of 

    nullity 1. However, from Lemma 3.4,  we can deduce that   A[T ∗]   is singular, which is a contradiction.

  • 8/17/2019 The Acyclic Matrices With a P Set of Maximum Size 2015 Linear Algebra and Its Applications

    8/11

    34   Z. Du, C.M. da Fonseca / Linear Algebra and its Applications 468 (2015) 27–37 

    (b)  ⇒  (a) Suppose that   T   is a tree obtained from   P 2   by sequentially adding pendant   P 2 ’s. Clearly,P s( A(P 2)) = 1. By Lemma 4.1, we have

    P s A(T )= P s A(P 2)+ n − 2

    2

      = n

    2

    .

    So (a) follows.  

    The example of a path of even order is perhaps the most simple application of   Theorem 4.2.

    Therefore, the Jacobi matrix of even order

    0 1

    1. . .

      . . .. . .

      . . .   1

    1 0

    in  [12, Theorem 4(a)] is an immediate consequence of our result.

    5. The odd case

    In order to establish our classification for the odd case, we need several technical lemmas.

    Lemma 5.1. (See [5, Corollary 3.5].) If T is a tree obtained from P 2  by sequentially adding pendant P 2’s, then

     for any vertex v in T , there exists an edge e incident with v such that T is a tree obtained from P 2 = e bysequentially adding pendant P 2’s.

    Lemma 5.2. Suppose that T is a tree on n 3 vertices, where n is odd, containing an outer P 3 . If there exists

    a matrix A ∈ S (T ) such that P s( A) =  n−1

    2   , then T is a tree obtained from P 3  by sequentially adding pendant 

    P 2 ’s.

    Proof.   Let   P 3 =  xz y  be an outer   P 3   in   T , where   x, y  are both terminal vertices. Denote by   z 1, . . . , z r the neighbors of   z   in   T   different from  x, y , and   T i , for 1 i r , the branch of   T   at   z  containing   z i .

    Suppose that   A   is a matrix in  S (T )  such that   P s( A) =  n−1

    2  , and  α   is a P-set of   A   with   |α| =   n−12   .

    From   Theorem 2.2(ii)–(iv), we know that each component of   T (α)   contains at most one edge.Furthermore, from Lemma 3.3,  exactly one of  x, y, z   is in  α.

    Let n i  be the number of vertices in  T i  and set  ñi  = |α ∩ V (T i )|, for 1 i r . Clearly,  n1 + · · · + nr  =n − 3, and  ñ1 + · · · + ñr  = |α| − 1 =

      n−32   .

    First suppose that   z  ∈ α . Note that   (α ∩  V (T i)) ∪ { z }   is a P-set of   A, equivalently,   α ∩  V (T 

    i)   is a

    P-set of   A( z ), for 1 i r . Since   T i   is a component of   T ( z ), we have  α ∩ V (T i )   is a P-set of   A[T i ], for1 i r .

    Suppose now that   z   /∈ α , and assume that   x ∈  α   and   y   /∈ α . Then   yz   is an edge of   T (α). ByTheorem 2.2(iv),  m A (0) = 0 and each component of   T (α)  contains at most one edge and of nullity 1.Moreover, by Lemma 3.5,  α ∩ V (T i )   is a P-set of   A[T i ], for 1 i r .

    Therefore, whether   z  ∈ α   or not,  α ∩  V (T i )   is always a P-set of   A[T i ], for 1 i r . Consequently,by Proposition 1.1,

    ñi P s A[T i]

    ni

    2

    ni

    2 ,

    for 1 i r , and thus

    n − 3

    2  = ñ1 + · · · + ñr 

    n1

    2  + · · · +

     nr 

    2  =

     n − 3

    2  .

    So, in fact, we have  ñi =  ni

    2 , furthermore,   P s( A[T i ]) =  ni

    2 , for 1 i r .

  • 8/17/2019 The Acyclic Matrices With a P Set of Maximum Size 2015 Linear Algebra and Its Applications

    9/11

     Z. Du, C.M. da Fonseca / Linear Algebra and its Applications 468 (2015) 27–37    35

    Clearly,   A[T i ] ∈ S (T i ), thus by Theorem 4.2 (a)  ⇒ (b),  T i  is a tree obtained from   P 2  by sequentiallyadding pendant   P 2’s, for 1 i r . Furthermore, it follows from Lemma 5.1 that there exists an edge

    ei   in   T i   incident with   z i   such that   T i   is a tree obtained from   P 2 =  e i   by sequentially adding pendantP 2 ’s, for 1 i r .

    Finally we can obtain   T   from   P 3 = xz y   by sequentially adding pendant   P 2’s as follows:

    () Firstly, we begin with the path   P 3 =  xz y, sequentially adding pendant   P 2 ’s, each time it wouldgive rise to two new edges:   zz i   and   ei , for 1 i   r , which finally would result in the tree

    obtained by attaching  r  paths   P 2  to the center   z  of the path   P 3 =  xz y.() Secondly, we begin with the resulting graph in (), sequentially adding pendant   P 2 ’s, just as

    the procedure of sequentially adding pendant   P 2 ’s from   ei   to   T i , for all 1 i r , which finally

    would result in the tree   T .

    Then the result follows.  

    Lemma 5.3. Suppose that T is a tree on n 3 vertices, where n is odd, containing a pendant P 3 =  uv w. Let T̄ be the tree obtained from T by deleting the vertices u, v. Suppose that A ∈ S (T ), and α  is a P-set of A with|α| =   n−12   . Let  β  = {u, v, w} ∩ α.

    (i)   If  β  = {u}, then P s( A[T̄ ]) =  n−3

    2   .

    (ii)   If  β  = {w}, then T is a tree obtained from P 3  by sequentially adding pendant P 2 ’s.(iii)   If  β  = {v} and A[u] = 0, then P s( A[T̄ ]) =

      n−32

      .

    (iv)   If  β  = {v} and A[u]     = 0, then T is a tree obtained from P 3 by sequentially adding pendant P 2 ’s.

    Proof.   Let  ᾱ = α ∩  V (T̄ ), i.e.,  ᾱ = α \ {u, v}, and let   ¯ A =  A[T̄ ]. Denote by   w1, . . . , w r   (all) the neigh-bors of   w   in   T   different from   v , and   T i , for 1 i r , the branch of   T   at   w  containing   w i . Finally, set

    αi  = α ∩ V (T i )  and   Ai  =  A[T i ], for 1 i r .(i) Suppose that  β = {u}. Then  v w  is an edge of  T (α), and from Theorem 2.2(iv),  { w1, . . . , wr } ⊆ α.Since   {u, w 1, . . . , wr } ⊆  α , thus   u   is a P-vertex of   A(w1, . . . , w r ), equivalently,   u   is a P-vertex of 

     A[u, v, w]. Now by Lemma 2.4(i), we can get that  m A[u,v,w](0) = m A[w](0).Note that  {w1, . . . , w r } ⊆ ᾱ. Clearly,

     A(ᾱ) =  A[u, v, w] ⊕ A 1(α1) ⊕ · · · ⊕  A r (αr )

    and

    ¯ A(ᾱ) =  A[w] ⊕ A 1(α1) ⊕ · · · ⊕  A r (αr ).

    So  m A(ᾱ)(

    0) =

    m¯ A(ᾱ)(

    0)

    .

    On the other hand, since  u   is a P-vertex of   A, by Lemma 2.4(i),  m A (0) = m ¯ A (0).We remark that  ᾱ ⊆ α   is a P-set of   A, i.e.,  m A(ᾱ)(0) = m A (0) + |ᾱ|, so  m ¯ A(ᾱ)(0) = m ¯ A (0) + |ᾱ|, i.e.,

    ᾱ   is a P-set of   ¯ A. Clearly,   |ᾱ| = |α| − 1 =   n−32   . Thus, by Proposition 1.1,   P s(¯ A) =   n−32   .

    (ii) Suppose that   β = {w}. Note that  αi  ∪ {w}, for 1 i r , is a P-set of   A, which implies that  αiis a P-set of   A(w). Since   T i   is a component of   T (w), we have  αi   is a P-set of   Ai , for 1 i r . Now,similarly to the proof of  Lemma 5.2, we can deduce that   T   is a tree obtained from   P 3  by sequentially

    adding pendant   P 2’s.

    (iii) Suppose that   β   = {v}   and   A[u] =  0. Note that   ᾱ  ⊆  α   is a P-set of   A, i.e.,   m A(ᾱ)(0)  =m A (0) + |ᾱ|. Since   A[u] = 0, from Lemma 2.4(ii), we can get  m A (0) = m ¯ A (0)  and m A(ᾱ)(0) = m ¯ A(ᾱ)(0).

    Now we have   m ¯ A(ᾱ)(0) = m ¯ A (0) + |ᾱ|, i.e.,  ᾱ   is a P-set of   ¯ A. Clearly,   |ᾱ| = |α| − 1 =   n−32   and, from

    Proposition 1.1,   P s( ¯ A) =   n−32   .(iv) Suppose that   β = {v}   and   A[u]     = 0. Taking into account that   u   is an isolated vertex of   T (α)

    and   A[u]     = 0, we have, from Theorem 2.2(iii),  m A (0) = 0,   A[w] = 0, and {w1, . . . , wr } ⊆ α.Since   v   is a P-vertex of   A, we have  m A(v)(0) = m A (0) + 1 = 1. From the fact that   u   is an isolated

    vertex of   T (v)   and   A[u]     =  0, it follows that   m A(u,v)(0) =  m A(v)(0) =  1. Additionally, from the fact

  • 8/17/2019 The Acyclic Matrices With a P Set of Maximum Size 2015 Linear Algebra and Its Applications

    10/11

    36   Z. Du, C.M. da Fonseca / Linear Algebra and its Applications 468 (2015) 27–37 

    that   v   is a P-vertex of   A   and  m A(u,v)(0)     = m A (0), we can deduce that  m A(v,w)(0) = m A (0) = 0, fromLemma 2.3. Thus   Ai  is nonsingular, for 1 i r .

    Now, since  {v, w 1, . . . , wr }   is a P-set of   A, thus   v  is a P-vertex of   A(w1, . . . , w r ), equivalently,   v   isa P-vertex of   A[u, v, w ], i.e.,  m A[u,w](0) = m A[u,v,w](0) + 1. Recall that   A[u]     = 0 and   A[w] =  0. Thus

    m A[u,w](0) = 1 and so   A[u, v, w ]  is nonsingular.Similarly to the proof of   Lemma   3.5, we have   αi   is a P-set of   Ai , for 1 i r . Moreover, analo-gously to the proof of  Lemma 5.2,   we can deduce that   T   is a tree obtained from   P 3   by sequentially

    adding pendant   P 2 ’s.  

    We are ready now to provide our second main classification.

     Theorem 5.4. Let T be a tree on n 3 vertices, where n is odd. The following two conditions are equivalent :

    (a)  There exists a matrix A  ∈ S (T ) such that P s( A) =  n−1

    2   .

    (b)  T is a tree obtained from P 3  by sequentially adding pendant P 2’s.

    Proof.   (a)  ⇒   (b) Suppose that   A   is a matrix in  S (T )  such that   P s( A) =  n−1

    2  .

    Let us start from   T , deleting pendant   P 2’s repeatedly, until there exists an outer   P 3 . This must

    occur as  n  is odd. The deletion process is listed as follows:

    T  = T 1 → ·· · → T k,

    where   T i+1   is the tree obtained from   T i  by deleting a pendant   P 2 , for 1 i k − 1.Let   Ai  =  A[T i ]  and  ni   be the order of   Ai , for 1 i k. Clearly,   A i ∈ S (T i ).By Proposition 1.1,   P s( Ai )

      ni −12

      , for 1 i k.

    Case 1.   P s( Ak) =  nk−1

    2   .

    Observe that there exists an outer   P 3   in   T k , from   Lemma 5.2,   T k   is a tree obtained from   P 3   by

    sequentially adding pendant   P 2’s, so   T   is a tree obtained from   P 3   by sequentially adding pendant

    P 2 ’s.

    Case 2.   P s( Ak) <  nk −1

    2  .

    Recall that   P s( A) =  n−1

    2   , i.e.,   P s( A1) =  n1−1

    2   , thus there exists an index, say   j, such that   P s( A j ) =n j −1

    2   and   P s( A j+1) <  n j+1−1

    2   .

    Let   α∗ be a P-set of   A j   with   |α∗| =

      n j −1

    2   . From  Theorem 2.2(ii)–(iv), each component of   T  j (α∗)

    contains at most one edge.Suppose that   T  j+1   is the tree obtained from   T  j   by deleting the pendant   P 2 =  u v , where   u   is a

    terminal vertex, and   w   is the neighbor of   v   in   T  j   different from  u.

    Note that   P s( A j+1) <  n j+1−1

    2  , then by Lemma 3.1,  exactly one of   u, v, w   is in  α∗. Moreover, from

    Lemma 5.3,  T  j  is a tree obtained from   P 3  by sequentially adding pendant   P 2’s. So  T  is a tree obtained

    from   P 3   by sequentially adding pendant   P 2 ’s.

    In conclusion, (a)  ⇒  (b).(b)  ⇒  (a) Suppose that   T   is a tree obtained from   P 3   by sequentially adding pendant   P 2 ’s. Clearly,

    P s( A(P 3)) = 1. By Lemma 4.1, we have

    P s A(T )= P s A(P 3)+ n − 3

    2

      = n − 1

    2

      .

    So (a) follows.  

    As in the even case, the path of odd order reveals to be the most simple example for the odd case.

    In fact, the matrices

  • 8/17/2019 The Acyclic Matrices With a P Set of Maximum Size 2015 Linear Algebra and Its Applications

    11/11

     Z. Du, C.M. da Fonseca / Linear Algebra and its Applications 468 (2015) 27–37    37

    0 1 01 0 1

    0 1 0

      and

    0 1 01 0 1

    0 1 1

    are, respectively, singular and nonsingular. Therefore, both Jacobi matrices of odd order

    0 1

    1. . .

      . . .

    . . .  . . .   1

    1 0

    and

    0 1

    1. . .

      . . .

    . . .  . . .   1

    1 0 1

    1 1

    attain the bound of  Proposition 1.1. The same is true if we append a path of even order to the central

    vertex of   P 3  which is the special type of the so-called  comet  considered by Kim and Shader in [12]:

    0 11

    . . .  . . .

    . . .  . . .   1

    1 0 1

    1 0 1 1

    1 0 0

    1 0 0

    and

    0 11

    . . .  . . .

    . . .  . . .   1

    1 0 1

    1 0 1 1

    1 0 0

    1 0 1

    .

    References

    [1]  M. An d̄elić, A. Erić, C.M. da Fonseca, Nonsingular acyclic matrices with full number of P-vertices, Linear Multilinear Algebra

    61 (1) (2013) 49–57.

    [2]   M. An  d̄elić, C.M. da Fonseca, R. Mamede, On the number of P-vertices of some graphs, Linear Algebra Appl. 434 (2) (2011)

    514–525.

    [3] Z. Du, C.M. da Fonseca, The singular acyclic matrices with the second largest number of P-vertices, submitted for publica-

    tion.

    [4]   Z. Du, C.M. da Fonseca, Nonsingular acyclic matrices with an extremal number of P-vertices, Linear Algebra Appl. 442

    (2014) 2–19.

    [5]   Z. Du, C.M. da Fonseca, The singular acyclic matrices with maximal number of P-vertices, Linear Algebra Appl. 438 (5)

    (2013) 2274–2279.

    [6]   A. Erić, C.M. da Fonseca, The maximum number of P-vertices of some nonsingular double star matrices, Discrete Math.

    313 (20) (2013) 2192–2194.

    [7]   C.M. da Fonseca, On the multiplicities of eigenvalues of a Hermitian matrix whose graph is a tree, Ann. Mat. Pura Appl.

    187 (2) (2008) 251–261.[8]  C.D. Godsil, Algebraic matching theory, Electron. J. Combin. 2 (1995) #R8.

    [9]  R.A. Horn, C.R. Johnson, Matrix Analysis, second edition, Cambridge University Press, New York, 2013.

    [10]  C.R. Johnson, A. Leal Duarte, C.M. Saiago, The Parter–Wiener theorem: refinement and generalization, SIAM J. Matrix Anal.

    Appl. 25 (2) (2003) 352–361.

    [11]  C.R. Johnson, B.D. Sutton, Hermitian matrices, eigenvalue multiplicities, and eigenvector components, SIAM J. Matrix Anal.

    Appl. 26 (2) (2004) 390–399.

    [12]  I.-J. Kim, B.L. Shader, Non-singular acyclic matrices, Linear Multilinear Algebra 57 (4) (2009) 399–407.

    http://refhub.elsevier.com/S0024-3795(13)00677-0/bib41454632303132s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib41454632303132s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib41454632303132s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib41454632303132s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib41454632303132s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib41454632303132s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib41464D32303131s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib41464D32303131s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib41464D32303131s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib41464D32303131s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib41464D32303131s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib41464D32303131s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib41464D32303131s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib44463230313362s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib44463230313362s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib44463230313361s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib44463230313361s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib454632303132s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib454632303132s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib454632303132s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib644632303038s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib644632303038s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib4731393935s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib484A31303133s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib4A445332303033s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib4A445332303033s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib4A5332303034s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib4A5332303034s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib4B5332303039s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib4B5332303039s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib4A5332303034s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib4A5332303034s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib4A445332303033s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib4A445332303033s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib484A31303133s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib4731393935s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib644632303038s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib644632303038s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib454632303132s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib454632303132s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib44463230313361s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib44463230313361s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib44463230313362s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib44463230313362s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib41464D32303131s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib41464D32303131s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib41454632303132s1http://refhub.elsevier.com/S0024-3795(13)00677-0/bib41454632303132s1