CHAPTER 1. MATRICES AND GAUSSIAN ELIMINATION Introduction What is Linear Algebra? ■ Linear Algebra deals with linear equations and linear functions which are represented through matrices and vectors. Ref: https://misa-xu.github.io/space/ 1 Why Study Linear Algebra? ■ Applications in communications – Multivariate signal processing and filter design – Multiple-input multiple-output (MIMO) system – Estimation and detection – Array signal processing ■ Applications in data science and machine learning – Vectorized code: Computation of many values simultaneously – Image recognition: Matrix representation of color image – Dimensionality reduction: Matrix fatorization – Data and learned model representation – Word embeddings 2 1.2 The Geometry of Linear Equation Geometry of Linear Equations (by rows) 2𝑥 − 𝑦 = 1 ■ Given a system with two equations and two unknowns ! 𝑥+𝑦 =5 ■ Look at the system by rows – Each equation is represented by a straight line in the 𝑥 − 𝑦 plane – The problem is to find the point of intersection of both lines 3 Geometry of Linear Equations (by columns) 2𝑥 − 𝑦 = 1 ■ Given a system with two equations and two unknowns ! 𝑥+𝑦 =5 ■ Look at the system by columns – Two separate equations are realized as one vector equation 2 −1 1 𝑥+ 𝑦= 1 1 5 – The problem is to find the combination of the column vectors on the LHS that produces the vector on the RHS LHS: left-hand side RHS: right-hand side 4 Geometry of Linear Equations (3-dim) ■ Given a system with three equations and three unknowns 2𝑢 + 𝑣 + 𝑤 = 5 * 4𝑢 − 6𝑣 = −2 −2𝑢 + 7𝑣 + 2𝑤 = 9 ■ Look at the system by rows ■ Look at the system by columns 2 1 1 5 4 𝑢 + −6 𝑣 + 0 𝑤 = −2 = 𝒃 −2 7 2 9 5 Geometry of Linear Equations (𝑛-dim) ■ With 𝑛 equations in 𝑛 unknowns, there are – Row picture: 𝑛 planes ■ Each equation produces (𝑛 − 1)-dimensional plane in 𝑛 dimensions – Column picture: 𝑛 vectors plus a vector 𝒃 on the RHS ■ The 𝑛 equations ask for (if there exists a unique solution) – Row picture: an intersection of 𝑛 planes ■ Every new plane (i.e., equation) reduces the dimension by one ■ When a point lies on all the planes, it is the solution! (i.e., the set of equations is said to be consistent) – Column picture: a linear combination of the 𝑛 columns that equals 𝒃 ■ For certain equations, it is impossible to find the above case, i.e., the singular case 6 The Singular Case (by rows) 𝑢 + 𝑣+ 𝑤=2 !2𝑢 + 3𝑤 = 5 3𝑢 + 𝑣 + 4𝑤 = 6 𝑢 + 𝑣+ 𝑤=2 !2𝑢 + 3𝑤 = 5 3𝑢 + 𝑣 + 4𝑤 = 7 ■ No solution as in Fig. 1.5(b) (inconsistent equations) ■ An infinity of solution as in Fig. 1.5(c) (end view: three planes have a whole line in common) ■ No solution as in Fig 1.5(d) 7 The Singular Case (by columns) ■ Three columns on the RHS are in the same plane, which is solvable only for 𝒃 in that plane 1 1 1 2 𝑢+ 0 𝑣+ 3 𝑤=𝒃 3 1 4 2 𝒃= 5 6 2 𝒃= 5 7 8 1.3 An Example of Gaussian Elimination Linear Equation and Gaussian Elimination ■ A basic problem in linear algebra is to solve a linear system 𝑎''𝑥' + 𝑎'(𝑥( + ⋯ + 𝑎') 𝑥) = 𝑏' 𝑎('𝑥' + 𝑎((𝑥( + ⋯ + 𝑎() 𝑥) = 𝑏( ⋮ 𝑎*'𝑥' + 𝑎*(𝑥( + ⋯ + 𝑎*) 𝑥) = 𝑏* – These are 𝑚 linear equation in 𝑛 unknowns 𝑥', … , 𝑥) . The 𝑎+, and 𝑏, are constant, and if 𝑏, = 0, 𝑗 = 1, … , 𝑚 the system is homogeneous. Otherwise, it is called inhomogeneous. ■ The locus of one equation 𝑎+'𝑥' + 𝑎+(𝑥( + ⋯ + 𝑎+) 𝑥) = 𝑏+ is an 𝑛 − 1 dimensional plane. – E.g., for 𝑎𝑥 + 𝑏𝑦 + 𝑐𝑧 = 𝑑, the locus of (𝑥, 𝑦, 𝑧) satisfying the equation is a plane in three-dimensional space ℝ- Locus: a set of all points whose location is determined by the equation 9 Example: Locus of Equation ■ For 𝑎𝑥 + 𝑏𝑦 + 𝑐𝑧 = 𝑑, the locus of (𝑥, 𝑦, 𝑧) satisfying the equation is a plane in three-dimensional space ℝ- 𝑎⃗𝚤 + 𝑏⃗𝚥 + 𝑐𝑘 is normal to the plane ■ For 𝑎𝑥 + 𝑏𝑦 = 𝑐, the locus is a line in ℝ( ! slope − " if b ≠ 0, otherwise vertical 10 Example: Locus of Equation ■ In parameter form, a line can be described by one parameter 𝑥! 𝑐! 𝑑! ⋮ =𝑡 ⋮ + ⋮ 𝑥" 𝑐" 𝑑" ■ Example 𝑥 =𝑡+1 – - 𝑦 = 2𝑡 𝑧 = 3𝑡 ■ A two-dimensional plane by two parameters 𝑥! 𝑐! 𝑒! 𝑑! ⋮ =𝑠 ⋮ +𝑡 ⋮ + ⋮ 𝑥" 𝑐" 𝑒" 𝑑" 11 Example: Locus of Equation (cont.) ■ Consider the plane in ℝ- given by 𝑥 + 2𝑦 + 3𝑧 = 1 *4𝑥 + 5𝑦 + 6𝑧 = 4 7𝑥 + 8𝑦 + 9𝑧 = 8 ⇒ 𝑥 + 2𝑦 + 3𝑧 = 1 * −3𝑦 − 6𝑧 = 0 −6𝑦 − 12𝑧 = 1 so that there are no points of intersection of the three planes. ■ If the “8” in the original set is changed to “7”, then we have 𝑥 + 2𝑦 + 3𝑧 = 1 * −3𝑦 − 6𝑧 = 0 −6𝑦 − 12𝑧 = 0 and the locus of intersection is the line: 𝑧 = 𝑡, 𝑦 = −2𝑡, 𝑥 =1+𝑡 ■ If the “9” in the original equation is changed, then the intersection is a single point 12 Example: Locus of Equation (cont.) ■ Gaussian elimination provides a systematic way of doing such analysis – These involve row operation and back substitution ■ Row operation: 1. Interchange two rows 2. Multiply a row by a nonzero number 3. Add a multiple of a row to another row 13 Example: Gaussian Elimination 2𝑢 + 𝑣 + 𝑤 = 5 ■ Original system: - 4𝑢 − 6𝑣 = −2 −2𝑢 + 7𝑣 + 2𝑤 = 9 ■ Forward elimination step pivot 𝟐 1 1 5 𝟐 1 1 𝟐 1 1 5 5 4 −6 0 −2 ⇒ 0 −𝟖 −2 −12 ⇒ 0 −𝟖 −2 −12 −2 7 2 9 0 8 3 0 0 𝟏 14 2 – With a full set of pivots 2, 8, and 1, there is only one solution ■ Backward elimination step (i.e., back-substitution) – 𝑤 = 2 at the third equation – 𝑣 = 1 by substituting 𝑤 = 2 into the second equation – 𝑢 = 1 by using 𝑣 = 1 and 𝑤 = 2 ■ By definition, pivots cannot be zero because we need to divide by them p. 13 14 Example: Gaussian Elimination pivot 1𝑥 + 2𝑦 + 3𝑧 = 1 *4𝑥 + 5𝑦 + 6𝑧 = 4 7𝑥 + 8𝑦 + 9𝑧 = 7 ⇒ pivot 𝑥 + 2𝑦 + 3𝑧 = 1 * −3𝑦 − 6𝑧 = 0 −6𝑦 − 12𝑧 = 0 ⇒ 𝑥 + 2𝑦 + 3𝑧 = 1 * −3𝑦 − 6𝑧 = 0 0 =0 ■ Back substitution: – 𝑧 is “free variable” 𝑧=𝑡 𝑦 = −2𝑡 – * 𝑥 = 1 + 4𝑡 − 3𝑡 = 1 + 𝑡 15 1.4 Matrix Notation and Matrix Multiplication Matrix Notation ■ For 𝑚×𝑛 matrix → → rows → 𝑎'' 𝑎(' ⋮ 𝑎*' ↑ ■ Column vector and row vector 𝑎' ⋮ and 𝑎* 𝑎'( 𝑎(( ⋮ 𝑎*( 𝑎') 𝑎(' ⋮ ⋯ 𝑎*) ↑ columns ↑ 𝑎' ⋯ ⋯ 𝑎+, ↑ 𝑖th row ↑ 𝑗th column ⋯ 𝑎) 16 Matrix Algebra ■ Addition of 𝑚×𝑛 matrices is point-wise 1 2 3 1 −1 2 2 1 5 – Ex. + = 4 5 6 −2 0 3 2 5 9 ■ Multiplication by scalar 1 2 3 2 4 6 – Ex. 2 = 4 5 6 8 10 12 ■ Inner product (dot product) of vectors 𝑏' 𝑎' ⋯ 𝑎) ⋅ ⋮ = 𝑎'𝑏'+ 𝑎(𝑏( + ⋯ + 𝑎) 𝑏) 𝑏) 17 Matrix-Vector Multiplication ■ Row-wise multiplication of 𝑚×𝑛 matrix and 𝑛 component column vector 𝒗' ⋅ 𝒘 𝒗' → 𝒘 𝒗( ⋅ 𝒘 𝒗( → ↓ = ← 𝑚 component column vector ⋮ ⋮ ⋮ 𝒗* ⋅ 𝒘 𝒗* → – 𝒗' ⋯ 𝒗* : 𝑛 component row vectors – 𝒘: 𝑛 component column vector ■ Example (row-wise) 1 2 3 1 −2 4 5 6 0 = −2 7 8 9 −1 −2 (row at a time) 18 Matrix-Vector Multiplication ■ Column-wise multiplication of 𝑚×𝑛 matrix and 𝑛 component column vector 𝒖' 𝒖( ⋯ 𝒖) 𝑤' ⋮ = 𝑤'𝒖' + 𝑤(𝒖( + ⋯ + 𝑤) 𝒖) ← 𝑚 component column vector ↓ ↓ ↓ 𝑤) – 𝒖' ⋯ 𝒖) : 𝑚 component column vectors – 𝒘: 𝑛 component column vector ■ Example (column-wise) 1 2 3 1 1 2 3 −2 4 5 6 0 = 1 4 + 0 5 + (−1) 6 = −2 7 8 9 −1 7 8 9 −2 19 Matrix-Matrix Multiplication ■ Multiplication of 𝑚×𝑛 matrix and 𝑛×𝑝 matrix (in that order) gives 𝑚×𝑝 matrix 𝒖! ⋅ 𝒗! 𝒖! ⋅ 𝒗# ⋯ 𝒖! ⋅ 𝒗% 𝒖! → 𝒗! 𝒗# ⋯ 𝒗% 𝒖# ⋅ 𝒗! 𝒖# ⋅ 𝒗# ⋯ 𝒖# ⋅ 𝒗% 𝒖# → ↓ ↓ ↓ = ⋮ ⋮ ⋮ ⋮ ⋮ 𝒖$ → 𝒖$ ⋅ 𝒗! 𝒖$ ⋅ 𝒗# ⋯ 𝒖$ ⋅ 𝒗% ■ Example 1 2 3 4 5 6 ■ Formal rule 𝑎!! ⋯ 𝑎!" ⋮ ⋮ 𝑎$! ⋯ 𝑎$" 𝑏!! ⋮ 𝑏"! ⋯ ⋯ −1 2 10 −4 4 0 = 22 −4 1 −2 𝑐!! 𝑏!% ⋮ = ⋮ 𝑐$! 𝑏"% ⋯ ⋯ 𝑐!% ⋮ 𝑐$% where 𝑐&' = ∑"()! 𝑎&( 𝑏(' 20 Property of Matrix Multiplication ■ If dimensions are appropriate so that the multiplication is defined: 𝑨𝑩 𝑪 = 𝑨(𝑩𝑪) Usually 𝑨𝑩 ≠ 𝑩𝑨 𝑨 𝑩 + 𝑪 = 𝑨𝑩 + 𝑨𝑪 𝑨 + 𝑩 𝑪 = 𝑨𝑪 + 𝑩𝑪 – Ex. 1 2 3 0 4 −1 1 −2 = 0 −4 1 0 and 3 −1 1 1 0 3 2 3 = 4 −1 4 −2 ■ You must be careful in doing matrix multiplication – Ex. 1 2 −1 −1 3 0 = 2 and 0 1 1 1 1 0 ■ 𝑛×𝑛 identity matrix 𝑰 = 0 1 ⋮ ⋱ 0 0 ■ If 𝑨 is 𝑛×𝑛, then 𝑨𝑰 = 𝑰𝑨 = 𝑨 2 −1 3 = 0 1 −2 0 2 −3 0 3 ⋯ 0 ⋱ 0 functions like number 1 ⋱ ⋮ ⋯ 1 21 Elementary Matrix (3×3 Case) ■ 𝑬89 matrix formed by interchanging 𝑖th and 𝑗th rows in 𝑰 0 1 0 – Ex. 𝑬:; = 1 0 0 0 0 1 ■ 𝑬8 (𝑐) matrix formed by multiplying 𝑖th row by nonzero number 𝑐 in 𝑰 1 0 0 – Ex. 𝑬; (4) = 0 4 0 0 0 1 ■ 𝑬89 (𝑐) matrix formed by replacing 𝑖th row in 𝑰 by 𝑖th row + 𝑐 times 𝑗th row 1 4 0 – Ex. 𝑬:; (4) = 0 1 0 0 0 1 ■ Each row operation in an arbitrary matrix can be accomplished by multiplying on the left by the corresponding elementary matrix – Ex. 𝑬:; 1 4 −1 6 2 0 7 3 0 8 1 4 2 = 0 0 9 4 1 0 0 1 0 −1 1 6 2 0 7 3 0 8 −3 4 2 = −1 6 9 2 0 7 3 12 0 2 8 9 22 1.5 Triangular Factors and Row Exchanges Model System of 𝑨𝒙 = 𝑩 ■ Multiplication of elementary matrices – 𝑬+, 𝑬+, = 𝑰 – 𝑬+ 𝑐 𝑬+ 1/𝑐 = 𝑰 – 𝑬+, (𝑐) 𝑬+, (−𝑐) = 𝑰 ■ We can write the system below 𝑎''𝑥' + ⋯ + 𝑎') 𝑥) = 𝑏' ⋮ 𝑎*'𝑥' + ⋯ + 𝑎*) 𝑥) = 𝑏* in matrix form 𝑨𝒙 = 𝒃, where 𝑎'' ⋯ 𝑎') ⋮ , 𝑨= ⋮ 𝑎*' ⋯ 𝑎*) 𝑥' 𝒙= ⋮ , 𝑥) 𝑏' and 𝒃 = ⋮ 𝑏* 23 Forward Elimination ■ Consider Gaussian elimination applied to an 𝑛×𝑛 matrix 𝑨 ■ By row operations, 𝑨 can be brought to upper triangular form 𝑎!! ∗ 𝑎## ⋱ 0 𝑎"" 1 2 3 *+, +-.*/01+2 1 2 3 – Ex. 4 5 6 0 −3 −6 7 8 9 0 0 0 – Ex. Using elementary matrices gives 1 2 3 1 2 3 𝑬3# −2 𝑬3! −7 𝑬#! (−4) 4 5 6 = 0 −3 −6 7 8 9 0 0 0 1 2 3 1 2 3 Thus, 4 5 6 = 𝑬#! (4) 𝑬3! 7 𝑬3# 2 0 −3 −6 7 8 9 0 0 0 24 LU Decomposition ■ Observation: In our algorithm, we began with top row, and worked our way down. Thus, in 𝑬-( −2 , 𝑬-' −7 , 𝑬('(−4) (and hence in 𝑬(' 4 , 𝑬-' 7 , 𝑬-( 2 ) no changes are made above the diagonal. Their product is therefore also lower triangular. 𝑳 = 𝑬(' 4 𝑬-' 7 𝑬-( 2 1 0 0 1 0 0 1 0 0 1 0 0 = 4 1 0 0 1 0 0 1 0 = 4 1 0 0 0 1 7 0 1 0 2 1 7 2 1 ■ Thus, if there are no row exchanges we may write 𝑨 = 𝑳𝑼 ↑ lower triangular with 1! s on diagonal ↑ upper triangular matrix ■ This is the LU decomposition Appendix: Product of lower triangular matrices 25 LDU Decomposition ■ From here, we can go to a more symmetric form. Suppose the diagonal entries in 𝑼 are not zero 𝑑' 𝑢'( ⋯ 𝑢') 1 0 𝑑( ⋮ 1 𝑨= ⋱ ⋱ ∗ 0 1 𝑑) ■ Then, we may write 1 𝑨= 0 1 ⋱ ∗ 1 𝑑' ⋮ 0 ⋯ 0 ⋱ ⋮ ⋯ 𝑑) 1 2!" 3! ⋯ 1 0 ⋱ 2!# 3! ⋮ 1 ■ This is the LDU decomposition ↑ ↑ lower triangular with upper triangular with 1! s on diagonal ↑ 1! s on diagonal diagonal matrix 26 Row Exchanges and Permutation Matrices ■ Suppose we need row exchanges. These are done by 𝑬+, matrices. A product 𝑷 of such matrices is called a permutation matrix. 1 1 1 𝑬$! 5( 𝑬"! 5' 1 1 1 𝑬"$ 1 1 1 – Ex. 𝑨 = 1 1 3 0 0 2 0 3 6 2 5 8 0 3 6 0 0 2 ⇔ 𝑬(-𝑬-' −2 𝑬(' −1 𝑨 = 𝑼 ■ We wish to write 𝑷𝑨 = 𝑳𝑼 where 𝑷 is a permutation matrix. This means that row exchange must be done first 1 1 1 𝑬$! 5' 𝑬"! 5( 1 1 1 𝑬(-𝑨 = 2 5 8 0 3 6 1 1 3 0 0 2 𝑬-' −1 𝑬(' −2 𝑬(-𝑨 = 𝑼 𝑬(-𝑨 = 𝑬(' 2 𝑬-' 1 𝑼 𝑳 𝑷 • Nonsingular 𝑨: There is a permutation matrix 𝑷 that reorders the rows of A to avoid zeros in the pivot positions. Then, 𝑨𝒙 = 𝒃 has a unique solution • Singular 𝑨: No 𝑷 can produce a full set of pivots (elimination fails). 27 1.6 Inverses and Transposes Inverses and Transposes ■ If 4! 𝑨 is an 𝑛×𝑛 matrix and 𝑩𝑨 = 𝑰, then 𝑩 is the inverse of 𝑨, denoted by 𝑩 = 𝑨 – Not all matrices have inverses. An inverse is impossible when 𝑨𝒙 is zero and 𝒙 is nonzero ■ Note 1. The inverse exists if and only if elimination produces 𝑛 pivots ■ Note 2. The matrix 𝑨 cannot have two different inverses. Suppose 𝑩𝑨 = 𝑰 and also 𝑨𝑪 = 𝑰. Then 𝑩 = 𝑪. ■ Note 3. If 𝑨 is invertible, the one and only solution to 𝑨𝒙 = 𝒃 is 𝒙 = 𝑨4! 𝒃 ■ Note 4. Suppose there is a nonzero vector 𝒙 such that 𝑨𝒙 = 𝟎. Then 𝑨 cannot have an inverse ■ Note 5. A 2×2 matrix is invertible if and only if 𝑎𝑑 − 𝑏𝑐 is not zero 1 𝑎 𝑏 4! 𝑑 −𝑏 = 𝑐 𝑐 𝑎𝑑 − 𝑏𝑐 −𝑐 𝑎 ■ Note 6. A diagonal matrix has an inverse provided no diagonal entries are zero ■ A product 𝑨𝑩 of invertible matrices is inverted by 𝑩4! 𝑨4! 28 Calculation of Inverse: Gauss-Jordan Method ■ To understand 𝑨5' better, suppose we use a sequence of row operation to bring 𝑨 to upper triangular form. Then, 𝑬𝑨 = 𝑻 – 𝑬 is the product of elementary matrices – Now, 𝑬 has an inverse as we saw in the previous slide – Thus, 𝑨 = 𝑬5'𝑻. Now if 𝑨 is nonsingular, then 𝑻 has nonzero pivots ■ Therefore, we can bring 𝑨 to 𝑰 by a sequence of row operations so h𝑨 = 𝑰 𝑬 h is the product of elementary matrices – 𝑬 h 5'𝑰 = 𝑬 h 5', and so – Hence, 𝑨 = 𝑬 h 5' ⇒ 𝑬 h = 𝑩𝑬 h 5'𝑬 h ⇒ 𝑬 h=𝑩 𝑰 = 𝑩𝑨 = 𝑩𝑬 29 Example: Gauss-Jordan Method 2 1 1 ■ For 𝑨 = 4 −6 0 , find 𝑨5' −2 7 2 − 30 Summary ■ If 𝑩𝑨 = 𝑰, then 𝑨𝑩 = 𝑰 as well and we write 𝑨5' = 𝑩 ■ In this case, 𝑩 = 𝑬: ⋯ 𝑬' product of elementary matrices so 𝑬: ⋯ 𝑬'𝑨 = 𝑰 5' 5' ■ 𝑨 = 𝑬'5' ⋯ 𝑬5' : 𝑰 = 𝑬' ⋯ 𝑬: – So, a nonsingular matrix is the product of elementary matrices ■ We can also see the familiar algorithm: If a sequence of row operations bring 𝑨 to 𝑰, the same sequence on 𝑰 gives 𝑨5' 𝑬: ⋯ 𝑬'𝑨 = 𝑰 ⇒ 𝑨5' = 𝑬: ⋯ 𝑬' ⇒ 𝑨5' = 𝑬: ⋯ 𝑬'𝑰 ■ If 𝑨 has an inverse, we say 𝑨 is invertible so invertible = nonsingular – 𝑨 is nonsingular ⇔ 𝑨 has an inverse 31 Transpose of Matrix ■ The transpose of 𝑨5 of matrix 𝑨 switches rows and columns: 1 4 5 1 2 3 – Ex. = 2 5 4 5 6 3 6 ■ An 𝑛×𝑛 matrix is symmetric if 𝑨5 = 𝑨 (𝑎'& = 𝑎&' for every 𝑖, 𝑗) ■ An 𝑛×𝑛 matrix is skew-symmetric if 𝑨5 = −𝑨 (𝑎'& = −𝑎&' for every 𝑖, 𝑗) ■ Some rules for matrix algebra: – 𝑨𝑩 5 = 𝑩5 𝑨5 – If 𝑨 is invertible, 𝑨4! 5 = 𝑨5 4! – If 𝑨 and 𝑩 are 𝑛×𝑛 invertible, 𝑨𝑩 4! = 𝑩4! 𝑨4! – Any square matrix 𝑨 ∈ ℝ"×" can uniquely be written as sum of a symmetric and a skew- symmetric matrix, known as Toeplitz decomposition: 𝑨= 1 1 𝑨 + 𝑨< + (𝑨 − 𝑨< ) 2 2 symmetric skew-symmetric 32 Appendix: Locus ■ Definition: Consider any set of equations in the variables 𝑥', … , 𝑥) . The locus in ℝ) of this set of equations is the set of vectors that satisfy all of the equations. The plural of a locus is loci. ■ Definition: A line in ℝ( is the locus of the equation 𝑎𝑥 + 𝑏𝑦 = 𝑐 for 𝑎, 𝑏, 𝑐 ∈ ℝ. In general, the line is affine. The line is linear if 𝑐 = 0. ■ Definition: A plane in ℝ- is the locus of the equation 𝑎𝑥 + 𝑏𝑦 + 𝑐𝑧 = 𝑑 for 𝑎, 𝑏, 𝑐, 𝑑 ∈ ℝ. In general, the line is affine. The plane is linear if 𝑑 = 0. ■ Definition: A hyperplane in ℝ) is the locus of the equation 𝑎'𝑥' + 𝑎(𝑥( + ⋯ + 𝑎) 𝑥) = 𝑐. It has dimension 𝑛 − 1. In general, its is affine. The hyperplane is linear if 𝑐 = 0. 33 Appendix: Product of Lower Triangular Matrices ■ Suppose 𝑨 = 𝑎&' and 𝑩 = 𝑏&' are 𝑛×𝑛 lower triangular matrices – By definition 𝑎&' = 𝑏&' = 0 if 𝑖 < 𝑗 ■ Let 𝑪 = 𝑐&' = 𝑨𝑩. Then, it is sufficient to show that 𝑐&' = 0 when 𝑖 < 𝑗 " 𝑐&' = ^ 𝑎&( 𝑏(' ()! ■ Case I: 𝑘 ≤ 𝑖 < 𝑗 – 𝑏(' = 0 for 𝑘, 𝑗 ∈ {1,2, … , 𝑛} with 𝑘 < 𝑗 ■ Case II: 𝑖 < 𝑘 – 𝑎&( = 0 for 𝑖, 𝑘 ∈ {1,2, … , 𝑛} with 𝑖 < 𝑘 ■ Hence, for 𝑖 < 𝑗, we have & " 𝑐&' = ^ 𝑎&( 𝑏(' + ^ 𝑎&( 𝑏(' = 0 ()! 𝑏!" = 0 in Case I ()&7! 𝑎#! = 0 in Case II 34 Appendix: Product of Upper Triangular Matrices ■ Suppose 𝑨 = 𝑎&' and 𝑩 = 𝑏&' are 𝑛×𝑛 lower triangular matrices – By definition 𝑎&' = 𝑏&' = 0 if 𝑖 > 𝑗 ■ Let 𝑪 = 𝑐&' = 𝑨𝑩. Then, it is sufficient to show that 𝑐&' = 0 when 𝑖 > 𝑗 " 𝑐&' = ^ 𝑎&( 𝑏(' ()! ■ Case I: 𝑖 > 𝑗 ≥ 𝑘 – 𝑎&( = 0 for 𝑖, 𝑘 ∈ {1,2, … , 𝑛} with 𝑖 > 𝑘 ■ Case II: 𝑘 > 𝑗 – 𝑏(' = 0 for 𝑘, 𝑗 ∈ {1,2, … , 𝑛} with 𝑗 > 𝑘 ■ Hence, for 𝑖 > 𝑗, we have ' " 𝑐&' = ^ 𝑎&( 𝑏(' + ^ 𝑎&( 𝑏(' = 0 ()! 𝑎!" = 0 in Case I ()'7! 𝑏!" = 0 in Case II 35