Download - Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Transcript
Page 1: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Graphs, Matrices and Generalized Inverse

Dr. Manjunatha Prasad [email protected], [email protected]

Professor of MathematicsDepartment of Statistics, PSPH, MAHE, Manipal

CoordinatorCenter for Advanced Research in Applied Mathematics and Statistics (CARAMS), MAHE, Manipal

December 10-21, 2018

Page 2: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Contents

1 Introduction

2 Some Properties of Incidence Matrices

3 Minors of incidence matrix

4 Moore–Penrose Inverse of Incidence Matrix

5 Properties of 0-1 Incidence Matrix

6 Generalized inverse of Laplacian Matrix

7 More results on generalized inverses and graphs

8 References

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 2 / 50

Page 3: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Introduction

Notationsℝ : set of real numbersℝ𝑚×𝑛 : set of real matrices of size 𝑚 × 𝑛ℝ𝑛 : set of real column vectors of order 𝑛𝟏 : (1,… , 1)𝑇{𝟏} : span of 𝟏

For 𝐴 ∈ ℝ𝑚×𝑛,𝐴𝑇 : transpose of 𝐴𝐴∗ : conjucate transpose of 𝐴(𝐴) : column space of 𝐴(𝐴) : (𝐴𝑇 )rank(𝐴) : rank of the matrix 𝐴𝐴𝑖 : 𝑖th row of 𝐴𝐴𝑗 : 𝑗th column of 𝐴

For any two subspaces 𝑆, 𝑇 of ℝ𝑛,𝑆 ⊕ 𝑇 : 𝑆 + 𝑇 with 𝑆 ∩ 𝑇 = {0}𝑆 ⟂ 𝑇 : 𝑆 and 𝑇 are orthogonal to each other𝑆 ⦹ 𝑇 : 𝑆 + 𝑇 with 𝑆 ⟂ 𝑇

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 3 / 50

Page 4: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Introduction

Preliminaries

Given an 𝑚 × 𝑛 matrix 𝐴 over a real field, consider the following matrixequations, called Penrose Equations;

(1) 𝐴𝑋𝐴 = 𝐴 (2) 𝑋𝐴𝑋 = 𝑋(3) (𝐴𝑋)𝑇 = 𝐴𝑋 (4) (𝑋𝐴)𝑇 = 𝑋𝐴.

A matrix 𝐺 satisfying the condition(s)∙ (1) → generalized inverse of 𝐴, denoted by 𝐴−

∙ (2) → outer inverse of 𝐴, denoted by 𝐴=

∙ (1) − (4) → Moore-Penrose inverse of 𝐴, denoted by 𝐴+

{𝐴−} = {𝐺 + (𝐼 − 𝐺𝐴)𝑈 + 𝑉 (𝐼 − 𝐴𝐺) ∶ 𝐺 is any g-inverse of 𝐴,𝑈 and 𝑉 are arbitrary matrices} (1)

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 4 / 50

Page 5: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Introduction

Preliminaries

Definition (Space Equivalence)For any two matrices 𝐴 and 𝐵,

∙ 𝐴 ≃𝐶𝑆

𝐵 if (𝐴) = (𝐵)∙ 𝐴 ≃

𝑅𝑆𝐵 if (𝐴) = (𝐵)

∙ 𝐴 ≃𝑆𝑃

𝐵 if (𝐴) = (𝐵) and (𝐴) = (𝐵)

LemmaFor any matrix 𝐴, 𝐴+ ≃

𝑆𝑃𝐴𝑇 .

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 5 / 50

Page 6: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Introduction

Preliminaries

LemmaFor given non-null matrices 𝐴, 𝐵 and 𝐶 in ℝ𝑚×𝑛, the following areequivalent.

(i) 𝐵𝐴−𝐶 is invariant under the choices of 𝐴−

(ii) (𝐶) ⊆ (𝐴) and (𝐵) ⊆ (𝐴).

Sketch of the proof.(𝑖) ⟹ (𝑖𝑖). Let 𝐺 ∈ {𝐴−}. For 𝐴− = 𝐺 + (𝐼 −𝐺𝐴)𝑈 + 𝑉 (𝐼 − 𝐴𝐺), where𝑈, 𝑉 are arbitrary, 𝐵𝐴−𝐶 = 𝐵𝐺𝐶 implies that𝐵(𝐼 − 𝐺𝐴)𝑈𝐶 = 0 = 𝐵𝑉 (𝐼 − 𝐴𝐺)𝐶 . Hence, 𝐵(𝐼 − 𝐺𝐴) = 0 and(𝐼 − 𝐴𝐺)𝐶 = 0, proving (𝑖𝑖).(𝑖𝑖) ⟹ (𝑖). If 𝐶 = 𝐴𝑋 and 𝐵 = 𝑌 𝐴 for some 𝑋 and 𝑌 , then

𝐵𝐴−𝐶 = 𝑌 𝐴𝐴−𝐴𝑋 = 𝑌 𝐴𝑋,

independent of 𝐴−. □K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 6 / 50

Page 7: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Introduction

Some matrices associated with graphs

Graph

IncidenceMatrix

AdjacencyMatrix

LaplacianMatrix

DistanceMatrix

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 7 / 50

Page 8: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Introduction

Incidence Matrix

Let 𝐺 be a graph with 𝑉 (𝐺) = {1,… , 𝑛} and 𝐸(𝐺) = {𝑒1,… , 𝑒𝑚}.For any directed graph 𝐺, the incidence matrix, denoted by 𝑄(𝐺), is the 𝑛 × 𝑚matrix defined as

(𝑄(𝐺))𝑖𝑗 =

⎧⎪⎨⎪⎩0 vertex 𝑖 and edge 𝑒𝑗 are not incident1 𝑒𝑗 originates at 𝑖−1 𝑒𝑗 terminates at 𝑖

.

The incidence matrix of the graph in Figure1.1 is,

𝑄 =

⎛⎜⎜⎜⎜⎝

𝑒1 𝑒2 𝑒3 𝑒4 𝑒5 𝑒61 −1 1 −1 0 0 02 1 0 0 −1 0 03 0 −1 0 0 1 04 0 0 1 0 0 −15 0 0 0 1 −1 1

⎞⎟⎟⎟⎟⎠

1

2 3 4

5

𝑒2𝑒1

𝑒5

𝑒3

𝑒4 𝑒6

Figure 1.1: Directed Graph

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 8 / 50

Page 9: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Introduction

Some important facts about 𝑄(𝐺)

For any graph 𝐺,∙ Always the column sum of 𝑄(𝐺) is zero.

Therefore, the rows are dependent.∙ Rank of 𝑄(𝐺) ≤ 𝑛 − 1.∙ Consider any left null vector 𝑥 of 𝑄(𝐺). Then,

𝑥𝑇𝑄(𝐺) = 0 ⟹ 𝑥𝑖 − 𝑥𝑗 = 0 if 𝑖 ∼ 𝑗.⟹ 𝑥𝑖 = 𝑥𝑗∀𝑖, 𝑗, if 𝐺 is connected

⎛⎜⎜⎜⎜⎝

𝑒1 𝑒2 𝑒3 𝑒41 −1 1 −1 02 1 0 0 −13 0 −1 0 04 0 0 1 05 0 0 0 1

⎞⎟⎟⎟⎟⎠Lemma

If 𝐺 is a connected graph on 𝑛 vertices, then rank 𝑄(𝐺) = 𝑛 − 1.

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 9 / 50

Page 10: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Introduction

Rank of 𝑄(𝐺)

LemmaFor any directed graph 𝐺 with 𝑘 components, the rank of 𝑄(𝐺) = 𝑛 − 𝑘.

Sketch of the proof.Let 𝐺1,… , 𝐺𝑘 be the connected components of 𝐺. Then, after a relabeling ofvertices(rows) and edges (columns) if necessary, we have

𝑄(𝐺) =

⎡⎢⎢⎢⎣𝑄(𝐺1) 0 … 0

0 𝑄(𝐺2) … 0⋮ ⋮ ⋱ ⋮0 0 … 𝑄(𝐺𝑘)

⎤⎥⎥⎥⎦ .

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 10 / 50

Page 11: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Introduction

Rank of 𝑄(𝐺) (contd.)

Since 𝐺𝑖 is connected, rank 𝑄(𝐺𝑖) is 𝑛𝑖 − 1, where 𝑛𝑖 is the number of verticesin 𝐺𝑖, 𝑖 = 1,… , 𝑘. It follows that

rank 𝑄(𝐺) = rank 𝑄(𝐺1) +⋯ + rank 𝑄(𝐺𝑘)= (𝑛1 − 1) +⋯ + (𝑛𝑘 − 1)= 𝑛1 +⋯ + 𝑛𝑘= 𝑛 − 𝑘

This completes the proof. □

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 11 / 50

Page 12: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Introduction

Adjacency Matrix

Let 𝐺 be a graph with 𝑉 (𝐺) = {1,… , 𝑛} and 𝐸(𝐺) = {𝑒1,… , 𝑒𝑚}. Theadjacency matrix of 𝐺, denoted by 𝐴(𝐺), is the 𝑛 × 𝑛 matrix defined asfollows.

(𝐴(𝐺))𝑖𝑗 =

{1 if vertex 𝑖 and vertex 𝑗 are adjacent0 otherwise

.

1

2 3 4

5

𝑒2𝑒1

𝑒5

𝑒3

𝑒4 𝑒6𝐴(𝐺) =

⎛⎜⎜⎜⎜⎜⎝

1 2 3 4 51 0 1 1 1 02 1 0 0 0 13 1 0 0 0 14 1 0 0 0 15 0 1 1 1 0

⎞⎟⎟⎟⎟⎟⎠K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 12 / 50

Page 13: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Introduction

0–1 Incidence Matrix

Let 𝐺 be a graph with 𝑉 (𝐺) = {1,… , 𝑛} and 𝐸(𝐺) = {𝑒1,… , 𝑒𝑚}. The(vertex-edge) incidence matrix of 𝐺, which we denote by 𝑀(𝐺), or simply by𝑀 , is the 𝑛 × 𝑚 matrix defined as follows.

(𝑀(𝐺))𝑖𝑗 =

{0 vertex 𝑖 and edge 𝑒𝑗 are not incident1 otherwise.

The 0–1 incidence matrix of the graph inFigure 1.2 is,

𝑀(𝐺) =

⎛⎜⎜⎜⎜⎝

𝑒1 𝑒2 𝑒3 𝑒4 𝑒5 𝑒61 1 1 1 0 0 02 1 0 0 1 0 03 0 1 0 0 1 04 0 0 1 0 0 15 0 0 0 1 1 1

⎞⎟⎟⎟⎟⎠

1

2 3 4

5

𝑒2𝑒1

𝑒5

𝑒3

𝑒4 𝑒6

Figure 1.2: Directed GraphK. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 13 / 50

Page 14: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Introduction

Degree Matrix

The degree matrix of a graph is a diagonalmatrix with the vertex degrees on its diag-onal.

𝐷𝑒𝑔(𝐺) =

⎛⎜⎜⎜⎜⎝

1 2 3 4 51 3 0 0 0 02 0 2 0 0 03 0 0 2 0 04 0 0 0 2 05 0 0 0 0 3

⎞⎟⎟⎟⎟⎠

1

2 3 4

5

𝑒2𝑒1

𝑒5

𝑒3

𝑒4 𝑒6

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 14 / 50

Page 15: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Introduction

Laplacian Matrix

The laplacian matrix of a given graph 𝐺 is, denoted by 𝐿(𝐺), is given by

(𝐿(𝐺))𝑖𝑗 =

⎧⎪⎨⎪⎩−1 𝑖 ≠ 𝑗, vertex 𝑖 and vertex 𝑗 are adjacent0 𝑖 ≠ 𝑗 vertex 𝑖 and vertex 𝑗 are not adjacent𝑑𝑖 i=j

.

𝐿(𝐺) =

⎛⎜⎜⎜⎜⎜⎜⎝

1 2 3 4 5 61 3 −1 0 −1 −1 02 −1 3 −1 0 −1 03 0 −1 2 0 −1 04 −1 0 0 2 −1 05 −1 −1 −1 −1 5 −16 0 0 0 0 −1 1

⎞⎟⎟⎟⎟⎟⎟⎠.

∙1 ∙2 ∙3

∙4 ∙5 ∙6

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 15 / 50

Page 16: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Introduction

Laplacian Matrix (contd.)

Facts about 𝐿(𝐺)

∙ 𝐿(𝐺) = 𝑄(𝐺)𝑄(𝐺)𝑇 , considering any arbitrary directions for the edges

∙ 𝐿(𝐺) = 𝐷𝑒𝑔(𝐺) − 𝐴(𝐺)

∙ rank(𝐿(𝐺)) = rank(𝑄(𝐺)) = 𝑛 − 𝑘, where 𝑘 is the number of component of thegraph 𝐺

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 16 / 50

Page 17: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Introduction

Some important results

ResultsFor any connected (directed) graph 𝐺 on 𝑛 vertices,♣ (𝑄(𝐺))⦹ {𝟏} = ℝ𝑛

♣ (𝐿(𝐺)) = (𝐿(𝐺)) and (𝐿(𝐺))⦹ {𝟏} = ℝ𝑛

For any graphs 𝐺1 and 𝐺2 on 𝑛 vertices,♣ 𝐿(𝐺1) ≃

𝑆𝑃𝐿(𝐺2)

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 17 / 50

Page 18: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Introduction

Distance Matrix

Distance 𝑑(𝑖, 𝑗) between the vertices 𝑖and 𝑗 of a connected graph 𝐺 is thelength of a shortest path from 𝑖 to 𝑗.

∙1 ∙2 ∙3

∙4 ∙5 ∙6

𝑑(1, 5) = 1.

Distance Matrix

The distance matrix of a connectedgraph, denoted by 𝐷(𝐺), is given by,

(𝐷(𝐺))𝑖𝑗 = 𝑑(𝑖, 𝑗).

⎛⎜⎜⎜⎜⎜⎜⎝

1 2 3 4 5 61 0 1 2 1 1 22 1 0 1 2 1 23 2 1 0 2 1 24 1 2 2 0 1 25 1 1 1 1 0 16 2 2 2 2 1 0

⎞⎟⎟⎟⎟⎟⎟⎠K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 18 / 50

Page 19: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Some Properties of Incidence Matrices

Page 20: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Some Properties of Incidence Matrices

Some Properties of Incidence Matrices

LemmaLet 𝐺 be a graph on 𝑛 vertices. Columns 𝑗1,… , 𝑗𝑘 of 𝑄(𝐺) are linearlyindependent if and only if the corresponding edges of 𝐺 induce an acyclicgraph.

A matrix is said to be totally unimodular if the determinant of any squaresubmatrix of the matrix is either 0 or ±1. It is easily proved by induction onthe order of the submatrix that 𝑄(𝐺) is totally unimodular as seen in the nextresult.

Lemma

Let 𝐺 be a graph with incidence matrix 𝑄(𝐺). Then 𝑄(𝐺) is totallyunimodular.

Sketch of the proof.

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 20 / 50

Page 21: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Some Properties of Incidence Matrices

Some Properties of Incidence Matrices(contd.)

Consider the statement that any 𝑘 × 𝑘 submatrix of 𝑄(𝐺) has determinant 0 or±1. We prove the statement by induction on 𝑘. Clearly the statement holds for𝑘 = 1, since each entry of 𝑄(𝐺) is either 0 or ±1. Assume the statement to betrue for 𝑘 − 1 and consider a 𝑘 × 𝑘 submatrix 𝐵 of 𝑄(𝐺). If each column of 𝐵has a 1 and a −1, then 𝑑𝑒𝑡𝐵 = 0. Also, if 𝐵 has a zero column, then𝑑𝑒𝑡𝐵 = 0. Now, suppose 𝐵 has a column with only one nonzero entry, whichmust be ±1. Expand the determinant of 𝐵 along that column and useinduction assumption to conclude that det B must be 0 or ±1. □

Lemma

Let 𝐺 be a tree on 𝑛 vertices. Then any submatrix of 𝑄(𝐺) of order 𝑛 − 1 isnonsingular.

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 21 / 50

Page 22: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Some Properties of Incidence Matrices

Some Properties of Incidence Matrices(contd.)

Incidence Vector of a Path Let 𝐺 be a graph with the vertex set𝑉 (𝐺) = {1, 2,… , 𝑛} and the edge set 𝐸(𝐺) = {𝑒1,… , 𝑒𝑚}. Given a path 𝑃 in𝐺, the incidence vector of 𝑃 is an 𝑚 × 1 vector defined as follows. The entriesof the vector are indexed by 𝐸(𝐺). If 𝑒𝑖 ∈ 𝐸(𝐺) then the 𝑖th element of thevector is 0 if the path does not contain 𝑒𝑖. If the path contains 𝑒𝑖 then the entryis 1 or −1, according as the direction of the path agrees or disagrees,respectively, with 𝑒𝑖.Let 𝐺 be a tree with the vertex set {1, 2,… , 𝑛}. We identify a vertex, say 𝑛, asthe root. The path matrix 𝑃𝑛 of 𝐺 (with reference to the root 𝑛) is defined asfollows. The 𝑗th column of 𝑃𝑛 is the incidence vector of the (unique) pathfrom vertex 𝑗 to 𝑛, 𝑗 = 1,… , 𝑛 − 1.

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 22 / 50

Page 23: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Some Properties of Incidence Matrices

Some Properties of Incidence Matrices(contd.)

Theorem

Let 𝐺 be a tree with the vertex set {1, 2,… , 𝑛}. Let 𝑄 be the incidence matrixof 𝐺 and let 𝑄𝑛 be the reduced incidence matrix obtained by deleting row 𝑛 of𝑄. Then 𝑄−1

𝑛 = 𝑃𝑛.

Definition (Unimodular matrix)A square integer matrix is called unimodular if its determinant is ±1.

Theorem

Let 𝐴 be an 𝑛 × 𝑛 integer matrix. Then 𝐴 is nonsingular and admits an integerinverse if and only if 𝐴 is unimodular.

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 23 / 50

Page 24: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Some Properties of Incidence Matrices

Some Properties of Incidence Matrices(contd.)

Theorem (Smith Normal Theorem)

Let 𝐴 be an 𝑚 × 𝑛 integer matrix. Then there exist unimodular matrices 𝑆 and𝑇 of order 𝑚 × 𝑚 and 𝑛 × 𝑛, respectively, such that

𝑆𝐴𝑇 =

⎡⎢⎢⎢⎢⎢⎢⎢⎣

𝑧1 0 ⋯ ⋯ 00 𝑧2 ⋯ ⋯ 0⋮ ⋱ ⋮0 ⋯ 𝑧𝑟 ⋯ 00 ⋯ 0 0⋮ ⋱ ⋮ ⋮0 ⋯ 0 0

⎤⎥⎥⎥⎥⎥⎥⎥⎦where 𝑧1,… , 𝑧𝑟 are positive integers (called the invariant factors of 𝐴) suchthat 𝑧𝑖 divides 𝑧𝑖+1, 𝑖 = 1, 2,… , 𝑟 − 1. Furthermore, 𝑧1⋯ 𝑧𝑖 = 𝑑𝑖, where 𝑑𝑖 isthe greatest common divisor of all 𝑖 × 𝑖 minors of 𝐴, 𝑖 = 1,… ,min{𝑚, 𝑛}.

Theorem

Let 𝐺 be a graph with vertex set 𝑉 (𝐺) = {1, 2,… , 𝑛} and edge set{𝑒1,… , 𝑒𝑚}. Suppose the edges 𝑒1,… , 𝑒𝑛−1 form a spanning tree of 𝐺. Let𝑄1 be the submatrix of 𝑄 formed by the rows 1,… , 𝑛 − 1 and the columns𝑒1,… , 𝑒𝑛−1. Let 𝑞 = 𝑚 − 𝑛 + 1. Partition 𝑄 as follows:

𝑄 =[

𝑄1 𝑄1𝑁−1𝑇𝑄1 −1𝑇𝑄1𝑁

].

Set𝐵 =

[𝑄−1

1 00 0

],

𝑆 =[𝑄−1

1 01𝑇 1

], 𝑇 =

[𝐼𝑛−1 −𝑁0 𝐼𝑞

],

𝐹 =[

𝑄1−1𝑇𝑄1

], 𝐻 =

[𝐼𝑛−1 𝑁

].

Then the following assertions hold:

(𝑖) 𝐵 is an integer reflexive g-inverse of 𝑄.(𝑖𝑖) 𝑆 and 𝑇 are unimodular matrices.

(𝑖𝑖𝑖) 𝑆𝑄𝑇 =[𝐼𝑛−1 00 0

]is the Smith normal form of 𝑄.

(𝑖𝑣) 𝑄 = 𝐹𝐻 is an integer rank factorization of 𝑄.

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 24 / 50

Page 25: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Moore–Penrose Inverse of IncidenceMatrix

Page 26: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Moore–Penrose Inverse of Incidence Matrix

Moore–Penrose Inverse of IncidenceMatrix

Lemma

If 𝐴 is an 𝑚 × 𝑛 matrix, then for an 𝑛 × 1 vector 𝑥, 𝐴𝑥 = 0 if and only if𝑥𝑇𝐴+ = 0.

Lemma

If 𝐺 is connected, then 𝐼 −𝑄𝑄+ = 1𝑛𝐽 .

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 26 / 50

Page 27: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Moore–Penrose Inverse of Incidence Matrix

Moore–Penrose Inverse of IncidenceMatrix (contd.)

Construction of Moore-Penrose Inverse : Let 𝐺 be a graph with𝑉 (𝐺) = {1, 2,… , 𝑛} and 𝐸(𝐺) = {𝑒1,… , 𝑒𝑚}. Suppose the edges 𝑒1,… , 𝑒𝑛−1form a spanning tree of 𝐺. Partition 𝑄 as follows:

𝑄 =[𝑈 𝑉

]where 𝑈 is 𝑛 × (𝑛 − 1) and 𝑉 is 𝑛 × (𝑚 − 𝑛 + 1). Note that, 𝑉 = 𝑈𝐷 for some𝐷, 𝑈 has left inverse and 𝑈

[𝐼 𝐷

]is a rank factorization of 𝑄 for some 𝐷.

Since 𝑟𝑎𝑛𝑘(𝑄) = 𝑟𝑎𝑛𝑘(𝑈 ) = 𝑛 − 1,

𝑄+ =[𝐼𝐷𝑇

](𝐼 +𝐷𝐷𝑇 )−1(𝑈𝑇𝑈 )−1𝑈𝑇 =

[𝑋

𝐷𝑇𝑋

]for 𝑋 = (𝐼 +𝐷𝐷𝑇 )−1(𝑈𝑇𝑈 )−1𝑈𝑇 .

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 27 / 50

Page 28: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Moore–Penrose Inverse of Incidence Matrix

Moore–Penrose Inverse of IncidenceMatrix (contd.)

Let 𝑀 = 𝑄𝑄+. Then

𝑀 = 𝑄𝑄+ = 𝑈[𝐼 𝐷

] [ 𝑋𝐷𝑇𝑋

]= 𝑈

[𝐼 +𝐷𝐷𝑇 ]𝑋.

Hence,𝐼 − 1

𝑛𝐽 = 𝑈

[𝐼 +𝐷𝐷𝑇 ]𝑋.

Hence, Thus, for any 𝑖, 𝑗,

𝑈𝑖(𝐼 +𝐷𝐷𝑇 )𝑋𝑗 = 𝑀(𝑖, 𝑗),

where 𝑈𝑖 is 𝑈 with row 𝑖 deleted, and 𝑋𝑗 is 𝑋 with column 𝑗 deleted.

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 28 / 50

Page 29: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Moore–Penrose Inverse of Incidence Matrix

Moore–Penrose Inverse of IncidenceMatrix (contd.)

By Lemma 8, 𝑈𝑖 is nonsingular. Also, 𝐷𝐷𝑇 is positive semidefinite and thus𝐼 +𝐷𝐷𝑇 is nonsingular. Therefore, 𝑈𝑖(𝐼 +𝐷𝐷𝑇 ) is nonsingular and

𝑋𝑗 = (𝑈𝑖(𝐼 +𝐷𝐷𝑇 ))−11𝑀(𝑖, 𝑗).

Once 𝑋𝑗 is determined, the 𝑗th column of 𝑋 is obtained using the fact that𝑄+1 = 0. Then 𝑌 is determined, since 𝑌 = 𝐷𝑇𝑋.

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 29 / 50

Page 30: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Moore–Penrose Inverse of Incidence Matrix

Moore–Penrose Inverse of IncidenceMatrix (contd.)

We illustrate the above method of calculating 𝑄+ by an example. Consider thegraph

1

2

3

4

𝑒1

𝑒4

𝑒2

𝑒5

𝑒3

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 30 / 50

Page 31: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Moore–Penrose Inverse of Incidence Matrix

Moore–Penrose Inverse of IncidenceMatrix (contd.)

with the incidence matrix

⎡⎢⎢⎢⎣1 0 0 1 0−1 1 0 0 10 −1 −1 0 00 0 1 −1 −1

⎤⎥⎥⎥⎦Fix the spanning tree formed by {𝑒1, 𝑒2, 𝑒3}. Then 𝑄 = 𝑈𝑉 where 𝑈 isformed by the first three columns of 𝑄. Observe that 𝑉 = 𝑈𝐷, where

𝐷 =⎡⎢⎢⎣1 01 1−1 −1

⎤⎥⎥⎦ .K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 31 / 50

Page 32: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Moore–Penrose Inverse of Incidence Matrix

Moore–Penrose Inverse of IncidenceMatrix (contd.)

Set 𝑖 = 𝑗 = 4. Then 𝑄+ =[𝑋𝑌

]where

𝑋4 = (𝑈4(𝐼 +𝐷𝐷𝑇 ))−1𝑀(4, 4) = 18

⎡⎢⎢⎣3 −2 −11 2 −31 0 −3

⎤⎥⎥⎦ .The last column of 𝑋 is found using the fact that the row sums of 𝑋 are zero.Then 𝑌 = 𝐷𝑇𝑋. After these calculations we see that

𝑄+ =[𝑋𝑌

]= 1

8

⎡⎢⎢⎢⎢⎢⎣

3 −2 −1 01 2 −3 01 0 −3 23 0 −1 −20 2 0 −2

⎤⎥⎥⎥⎥⎥⎦.

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 32 / 50

Page 33: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Properties of 0-1 Incidence Matrix

Page 34: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Properties of 0-1 Incidence Matrix

Properties of 0-1 Incidence Matrix

Lemma

Let 𝐶𝑛 be the cycle on the vertices {1,… , 𝑛}, 𝑛 ≥ 3, and let 𝑀 be itsincidence matrix. Then 𝑑𝑒𝑡 𝑀 equals 0 if 𝑛 is even and ±2 if 𝑛 is odd.

Lemma

Let 𝐺 be a connected graph with 𝑛 vertices and let 𝑀 be the incidence matrixof 𝐺. Then the rank of 𝑀 is 𝑛 − 1 if 𝐺 is bipartite and 𝑛 otherwise.

Sketch of the proof.Suppose 𝐺 is not bipartite graph, 𝐺 has a cycle of odd length. To prove𝑟𝑎𝑛𝑘(𝑀) = 𝑛, consider 𝑥𝑇𝑀 = 0. Then 𝑥𝑖 + 𝑥𝑗 = 0 whenever the vertices 𝑖and 𝑗 are adjacent. Since 𝐺 is connected it follows that |𝑥𝑖| = 𝛼, 𝑖 = 1,… , 𝑛,for some constant 𝛼. Suppose 𝐺 has an odd cycle formed by the vertices

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 34 / 50

Page 35: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Properties of 0-1 Incidence Matrix

Properties of 0-1 Incidence Matrix (contd.)

𝑖1,… , 𝑖𝑘. Suppose 𝑥𝑖1 = 𝛼, then 𝑥𝑖2 = −𝛼, 𝑥𝑖3 = 𝛼,… 𝑥𝑖𝑘 = 𝛼, which impliesthat 𝑥𝑖𝑘+1 = 𝑥𝑖1 = −𝛼. That is 𝛼 = −𝛼, implies that 𝛼 = 0. Hence, 𝑥 = 0which proves rank of 𝑀 is 𝑛.Now suppose 𝐺 has no odd cycle, that is, 𝐺 is bipartite. Let 𝑉 (𝐺) = 𝑋 ∪ 𝑌 bea bipartition. Orient each edge of 𝐺 as by considering all the edges orginatesfrom 𝑌 and let 𝑄 be the corresponding {0, 1,−1}–incidence matrix. We knowthat 𝑟𝑎𝑛𝑘(𝑄) = 𝑛− 1. Also we know that 𝑄 = 𝐷𝑀 where 𝐷 is 𝑛 × 𝑛 diagonalmatrix whose 𝑖th entry is −1 if 𝑖 ∈ 𝑌 , otherwise 1.Clearly 𝐷 is invertible and therefore,

𝑛 − 1 = 𝑟𝑎𝑛𝑘 𝑄 = 𝑟𝑎𝑛𝑘 𝐷𝑀 = 𝑟𝑎𝑛𝑘 𝑀.

Hence, the proof.□

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 35 / 50

Page 36: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Generalized inverse of LaplacianMatrix

Page 37: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Generalized inverse of Laplacian Matrix

Generalized inverse of Laplacian Matrix

LemmaFor a tree 𝑇 on 𝑛 vertices, let 𝐷 be the distance matrix, 𝐿 the Laplacianmatrix and 𝜏 is a column vector with 𝜏𝑖 = 2 − 𝑑𝑖, where 𝑑𝑖 is the degree of the𝑖th vertex. Then,

𝐿𝐷 + 2𝐼 = 𝜏𝟏𝑇 .

Sketch of the proof.Note that,

𝐿 =

⎡⎢⎢⎢⎣𝑑1 −𝑎12 … −𝑎1𝑛

−𝑎21 𝑑2 … −𝑎2𝑛⋮ ⋮ ⋱ ⋮

−𝑎𝑛1 −𝑎𝑛2 … 𝑑𝑛

⎤⎥⎥⎥⎦ and 𝐷 =

⎡⎢⎢⎢⎣0 𝑑(1, 2) … 𝑑(1, 𝑛)

𝑑(2, 1) 0 … 𝑑(2, 𝑛)⋮ ⋮ ⋱ ⋮

𝑑(𝑛, 1) 𝑑(𝑛, 2) … 0

⎤⎥⎥⎥⎦.

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 37 / 50

Page 38: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Generalized inverse of Laplacian Matrix

Generalized inverse of Laplacian Matrix(contd.)To find, (𝐿𝐷)𝑖𝑗 , without loss of generality, assume that 𝑑𝑖 = 𝑘 and{1, 2,… , 𝑘} are adjacent with 𝑖.For 𝑗 = 𝑖,

(𝐿𝐷)𝑖𝑖 = 𝐿𝑖𝐷𝑖

= −𝑎𝑖1𝑑(𝑖, 1) − 𝑎𝑖2𝑑(𝑖, 2) −⋯ − 𝑎𝑖𝑘𝑑(𝑖, 𝑘)= −𝑘 = −𝑑𝑖 = 𝜏𝑖 − 2.

For 𝑗 ≠ 𝑖,

(𝐿𝐷)𝑖𝑗 = 𝐿𝑖𝐷𝑗

= 𝑙𝑖𝑖𝑑(𝑖, 𝑗) −∑𝑝≠𝑖

𝑎𝑖𝑝𝑑(𝑝, 𝑗)

= 𝑘𝑑(𝑖, 𝑗) − (𝑘(𝑑(𝑖, 𝑗) + 1) − 2)= 2 − 𝑘 = 2 − 𝑑𝑖 = 𝜏𝑖

∙𝑖

∙1 ∙2 ∙𝑘

∙𝑗

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 38 / 50

Page 39: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Generalized inverse of Laplacian Matrix

Generalized inverse of Laplacian Matrix(contd.)

Therefore,𝐿𝐷 = 𝜏𝟏𝑇 − 2𝐼.

In other words, 𝐿𝐷 + 2𝐼 = 𝜏𝟏𝑇 .□

TheoremFor any tree 𝑇 and its Laplacian matrix 𝐿, a generalized inverse of 𝐿 is givenby −1

2𝐷, where 𝐷 is the distance matrix of 𝑇 . That is,

𝐿(−12𝐷)𝐿 = 𝐿.

Sketch of the proof.Proof follows from the above lemma and the fact that (𝐿) = {𝟏}⟂. □

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 39 / 50

Page 40: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Generalized inverse of Laplacian Matrix

For 𝑖 ≠ 𝑗, define

𝑒𝑖𝑗 =

⎛⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎝

0⋮010⋮0−10⋮0

⎞⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎠

.

𝑖

𝑗

For any square matrix 𝑋,

𝑒𝑇𝑖𝑗𝑋𝑒𝑖𝑗 = 𝑥𝑖𝑖 + 𝑥𝑗𝑗 − 𝑥𝑖𝑗 − 𝑥𝑗𝑖.

ObservationFor any connected graph 𝐺 on 𝑛vertices,

𝑒𝑖𝑗 ∈ (𝑄(𝐺)) = (𝐿(𝐺)) for all 𝑖 ≠ 𝑗

as 𝑒𝑇𝑖𝑗𝟏 = 0 and {𝟏}⟂ = (𝑄(𝐺)).

The following is immediate from the observation.

LemmaFor any connected graph 𝐺 and its Laplacian matrix 𝐿,

𝑒𝑇𝑖𝑗𝐿−𝑒𝑖𝑗 is invariant under the choice of 𝐿−.

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 40 / 50

Page 41: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Generalized inverse of Laplacian Matrix

Weiner Index of a graph

DefinitionFor any graph 𝐺, Weiner index of 𝐺, denoted by 𝑊 (𝐺), is defined by

𝑊 (𝐺) =∑𝑖<𝑗

𝑑(𝑖, 𝑗). (2)

TheoremIf 𝑇 is a tree with Laplacian matrix 𝐿 then

𝑊 (𝑇 ) = 𝑛𝑛−1∑𝑖=1

1𝜆𝑖, where 𝜆1 ≥ 𝜆2 ≥ … 𝜆𝑛−1 > 𝜆𝑛 = 0 are eigenvalues of 𝐿.

(3)

Sketch of the proof.

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 41 / 50

Page 42: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Generalized inverse of Laplacian Matrix

Weiner Index of a graph (contd.)

Note that, if 𝜆1 ≥ 𝜆2 ≥ … 𝜆𝑛−1 > 𝜆𝑛 = 0 are eigenvalues of 𝐿, then1𝜆1, 1𝜆2… 1

𝜆𝑛−1are the eigenvalues of 𝐿+ and

trace(𝐿+) =𝑛−1∑𝑖=1

1𝜆𝑖.

Now for any 𝑖, 𝑗, 𝑖 ≠ 𝑗,

𝑒𝑇𝑖𝑗(−12𝐷)𝑒𝑖𝑗 = −1

2(𝑑(𝑖, 𝑖) + 𝑑(𝑗, 𝑗) − 𝑑(𝑖, 𝑗) − 𝑑(𝑗, 𝑖)

)= 𝑑(𝑖, 𝑗).

Since 𝑒𝑇𝑖𝑗𝐿−𝑒𝑖𝑗 is invariance under choice of 𝐿−,

𝑑(𝑖, 𝑗) = 𝑒𝑇𝑖𝑗𝐿−𝑒𝑖𝑗 = 𝑒𝑇𝑖𝑗𝐿

+𝑒𝑖𝑗= 𝑙+𝑖𝑖 + 𝑙+𝑗𝑗 − 2𝑙+𝑖𝑗 ( Since 𝐿+ is symmetric)

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 42 / 50

Page 43: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Generalized inverse of Laplacian Matrix

Weiner Index of a graph (contd.)

∑𝑖,𝑗

𝑑(𝑖, 𝑗) =∑𝑖,𝑗

𝑙+𝑖𝑖 +∑𝑖,𝑗

𝑙+𝑗𝑗 − 2∑𝑖,𝑗

𝑙+𝑖𝑗

= 2𝑛 trace(𝐿+) (since 𝐿+𝟏 = 0)

= 2𝑛𝑛−1∑𝑖=1

1𝜆𝑖

Therefore,

𝑊 (𝑇 ) =∑𝑖<𝑗

𝑑(𝑖, 𝑗) = 𝑛𝑛−1∑𝑖=1

1𝜆𝑖

= 𝑛 trace(𝐿+).

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 43 / 50

Page 44: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

More results on generalized inversesand graphs

Page 45: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

More results on generalized inverses and graphs

Absorption Law

Let 𝑎, 𝑏 be two elements in an associative ring 𝑅 with outer inverses 𝑎= and 𝑏=respectively.

Absorption Law (Bapat, Jain, Karantha, and Raj[1])Absorption law (Extended absorption law) for the elements 𝑎 and 𝑏 in anassociative ring 𝑅 with reference to the outer inverses 𝑎= and 𝑏= is

𝑎=(𝑎 + 𝑏)𝑏= = 𝑎= + 𝑏=.

Lemma (Bapat, Jain, Karantha, and Raj[1]))𝑎=(𝑎 + 𝑏)𝑏= = 𝑎= + 𝑏= if and only if 𝑎=𝑅 ⊇ 𝑏=𝑅 and 𝑅𝑎= ⊆ 𝑏=𝑅.

If 𝐴 and 𝐵 are two matrices,

𝐴=(𝐴+𝐵)𝐵= = 𝐴=+𝐵= if and only if (𝐴=) = (𝐵=) and (𝐴=) = (𝐵=).

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 45 / 50

Page 46: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

More results on generalized inverses and graphs

Absorption Law for incidence matrices

TheoremLet 𝐺1 and 𝐺1 be any two connected graphs on 𝑛 vertices, and with 𝑚 edges.Let 𝑄1 = 𝑄(𝐺1) and 𝑄2 = 𝑄(𝐺2). Then the following are equivalent

(i) Absorption law ‘𝑄+1 (𝑄1 +𝑄2)𝑄+

2 = 𝑄+1 +𝑄+

2 ’ holds(ii) (𝑄+

1 ) = (𝑄+2 )

(iii) (𝑄1) = (𝑄2)

Sketch of the proof.Since for any connected graph 𝐺, (𝑄(𝐺))⦹ {𝟏} = ℝ𝑛,

(𝑄1) = (𝑄2) = {𝟏}⟂.

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 46 / 50

Page 47: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

More results on generalized inverses and graphs

Absorption Law for incidence matrices(contd.)

Therefore, from the necessary sufficient condition for the absorption law to besatisfied, (𝑖) holds if and only if

(𝑄1) ⊆ (𝑄2).

Equality holds from the fact that (𝑄1) = (𝑄2). Hence, (𝑖) ⟺ (𝑖𝑖𝑖).(𝑖𝑖) ⟺ (𝑖𝑖𝑖) follows from the fact that 𝐴 ≃

𝑆𝑃𝐴+. □

CorollaryLet 𝐺1 and 𝐺2 are two trees on 𝑛 vertices. Then absorption law given in theabove lemma always holds.

Sketch of the proof.Proof follows from the previous Theorem as (𝑄1) = (𝑄2) = ℝ𝑛−1. □

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 47 / 50

Page 48: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

More results on generalized inverses and graphs

Absorption Law for incidence matrices(contd.)

CorollaryLet 𝐺1 and 𝐺2 be any two connected graphs on 𝑛 vertices, and let 𝐿1 and 𝐿2be its laplacian matrices respectively. Then the absorption law

𝐿+1 (𝐿1 + 𝐿2)𝐿+

2 = 𝐿+1 + 𝐿+

2

Sketch of the proof.Proof follows from the above Theorem as (𝐿(𝐺)) = (𝐿(𝐺)) = {𝟏}⟂, forany connected graph 𝐺. □

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 48 / 50

Page 49: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

References

References

✓ Ravindra B. Bapat, Surender Kumar Jain, K. Manjunatha Prasad Karantha, andM. David Raj. “Outer inverses: Characterization and applications”. In: LinearAlgebra Appl. 528 (Sept. 2017), pp. 171–184.

✓ Adi Ben-Israel and T. N. E. Greville. Generalized inverses: Theory andapplications. Second. Springer-Verlag, Berlin, 2002.

✓ C. R. Rao and S. K. Mitra. Generalized inverses of matrices and applications.Wiley, New York, 1971.

✓ Ravindra B Bapat. Graphs and matrices. Springer, 2010.

K. Manjunatha Prasad (MAHE, Manipal) G-inverses & Graphs 06 December, 2018 49 / 50

Page 50: Graphs, Matrices and Generalized Inverse › ... › Day5_Session_1_Lecture_KMPK.pdf · Contents 1 Introduction 2 Some Properties of Incidence Matrices 3 Minors of incidence matrix

Thank You