Matrices and Linear Algebra

M. Matrices and Linear Algebra 1. Matrix algebra. In section D we calculated the determinants of square arrays of numbers. Such arrays are important in mathematics and its applications; they are called matrices. In general, they need not be square, only rectangular. A rectangular array of numbers having m rows and n columns is called an m x n matrix. The number in the i-th row and j-th column (where 1 5 i 5 m , 1 5 j 5 n) is called the ij-entry, and denoted aij; the matrix itself is denoted by A, or sometimes by (aij). Two matrices of the same size are equal if corresponding entries are equal. Two special kinds of matrices are the row-vectors: the 1 x n matrices ( a l , az, . . . ,a,); and the column vectors: the m x 1 matrices consisting of a column of m numbers. From now on, row-vectors or column-vectors will be indicated by boldface small letters; when writing them by hand, put an arrow over the symbol. Matrix operations There are four basic operations which produce new matrices from old. 1. Scalar multiplication: Multiply each entry by c : cA = (caij) + + 2. Matrix addition: Add the corresponding entries: A B = (aij bij); the two matrices must have the same number of rows and the same number of columns. 3. Transposition: The transpose of the m x n matrix A is the n x m matrix obtained by making the rows of A the columns of the new matrix. Common notations for the transpose are AT and A'; using the first we can write its definition as AT = (aji). If the matrix A is square, you can think of AT as the matrix obtained by flipping A over around its main diagonal. 2 -3 Example 1.1 Let A = 1 5 . F i n d A + B , AT, 2A-3B. 18.02 NOTES 2 4. Matrix multiplication This is the most important operation. Schematically, we have mxn nxp mxP The essential points are: 1. For the multiplication to be defined, A must have as many columns as B has rows; 2. The ij-th entry of the product matrix C is the dot product of the i-th row of A with the j-th column of B . The two most important types of multiplication, for multivariable calculus and differential equations, are: 1. AB, where A and B are two square matrices of the same size - these can always be multiplied; 2. Ab, where A is a square n x n matrix, and b is a column n-vector. Laws and properties of matrix multiplication M-1. A(B + C ) = AB + AC, M-2. (AB)C = A(BC); (A + B ) C = AC + B C (cA)B = c(AB). distributive laws associative laws In both cases, the matrices must have compatible dimensions. M-3. Let I3= ; then A I = A and I A = A for any 3 x 3 matrix. I is called the identity matrix of order 3. There is an analogously defined square identity matrix Inof any order n, obeying the same multiplication laws. M-4. In general, for two square n x n matrices A and B , AB # BA: matrix multiplication is not commutative. (There are a few important exceptions, but they are very special - for example, the equality A I = I A where I is the identity matrix.) M-5. For two square n x n matrices A and B , we have the determinant law: lABl = IAIJBI, also written det(AB) = det(A)det(B) For 2 x 2 matrices, this can be verified by direct calculation, but this naive method is unsuitable for larger matrices; it's better to use some theory. We will simply assume it in these notes; we will also assume the other results above (of which only the associative law M-2 offers any difficulty in the proof). M. MATRICES AND LINEAR ALGEBRA 3 M-6. A useful fact is this: matrix multiplication can be used to pick out a row or column of a given matrix: you multiply by a simple row or column vector to do this. Two examples should give the idea: (i % g ) (8) (i) = (1 0 0 ) (: : :) 4 5 6 = ( 1 the second column 2 3) thefirstrow Exercises: Section 1F 2. Solving square systems of linear equations; inverse matrices. Linear algebra is essentially about solving systems of linear equations, an important application of mathematics to real-world problems in engineering, business, and science, especially the social sciences. Here we will just stick to the most important case, where the system is square, i.e., there are as many variables as there are equations. In low dimensions such systems look as follows (we give a 2 x 2 system and a 3 x 3 system): In these systems, the aij and bi are given, and we want to solve for the xi. As a simple mathematical example, consider the linear change of coordinates given by the equations If we know the y-coordinates of a point, then these equations tell us its x-coordinates immediately. But if instead we are given the x-coordinates, to find the y-coordinates we must solve a system of equations like (7) above, with the yi as the unknowns. Using matrix multiplication, we can abbreviate the system on the right in (7) by where A is the square matrix of coefficients ( a i j ). (The 2 x 2 system and the n x n system would be written analogously; all of them are abbreviated by the same equation Ax = b, notice.) You have had experience with solving small systems like (7) by elimination: multiplying the equations by constants and subtracting them from each other, the purpose being to 18.02 NOTES 4 eliminate all the variables but one. When elimination is done systematically, it is an efficient method. Here however we want to talk about another method more compatible with handheld calculators and MatLab, and which leads more rapidly to certain key ideas and results in linear algebra. Inverse matrices. Referring to the system (8), suppose we can find a square matrix M , the same size as A, such that MA = I (9) (the identity matrix). We can then solve (8) by matrix multiplication, using the successive steps, where the step M(Ax) = x is justified by M ( A x ) = (MA)x, =Ix, = x, by M-2; by (9); by M-3 . Moreover, the solution is unique, since (10) gives an explicit formula for it. The same procedure solves the problem of determining the inverse to the linear change of coordinates x = Ay, as the next example illustrates. 1 2 2, Verify that M satisfies (9) ( 2 3) and M = (-32 -1 ' above, and use it to solve the first system below for xi and the second for the yi in terms of the xi: E x a m p l e 2.1 Let A = Solution. 1 2 We have ( 2 3) (-3 2 solve the first system, we have by (lo), (i :), ( ) ( -12, = = 2 by matrix multiplication. To -1 (3 (A:), = so the solution is x1 = 11,x2 = -6. By reasoning similar to that used above in going from Ax = b to x = M b , the solution to x = Ay is y = M x , so that we get as the expression for the yi in terms of the xi. Our problem now is: how do we get the matrix M ? In practice, you mostly press a key on the calculator, or type a Matlab command. But we need to be able to work abstractly M. MATRICES AND LINEAR ALGEBRA 5 with the matrix - i.e., with symbols, not just numbers, and for this some theoretical ideas are important. The first is that M doesn't always exist. M exists The implication IAI # 0. + follows immediately from the law M-5, since The implication in the other direction requires more; for the low-dimensional cases, we will produce a formula for M. Let's go to the formal definition first, and give M its proper name, A-l: Definition. Let A be an n x n matrix, with IAl matrix, written A-', such that # 0. Then the inverse of A is an n x n (It is actually enough to verify either equation; the other follows automatically - see the exercises.) Using the above notation, our previous reasoning (9) - (10) shows that (12) (12) + + #0 JAl # 0 IAl the unique solution of A x = b is x = A-'b; the solution of x = A y for the yi is y = A-'x. Calculating the inverse of a 3 x 3 matrix Let A be the matrix. The formulas for its inverse A-' and for an auxiliary matrix adj A called the adjoint of A (or in some books the adjugate of A) are T (13) A-' All = - adj A = - Azl I A1 1 1( A 1 1 A12 A13 Azz A t::) ' In the formula, Aij is the cofactor of the element aij in the matrix, i.e., its minor with its sign changed by the checkerboard rule (see section 1 on determinants). Formula (13) shows that the steps in calculating the inverse matrix are: 1. Calculate the matrix of minors. 2. Change the signs of the entries according to the checkerboard rule. 3. 3-anspose the resulting matrix; this gives adj A. 4. Divide every entry by IAl. (If inconvenient, for example if it would produce a matrix having fractions for every entry, you can just leave the 1/IA( factor outside, as in the formula. Note that step 4 can only be taken if IAl # 0, so if you haven't checked this before, you'll be reminded of it now.) The notation Aij for a cofactor makes it look like a matrix, rather than a signed determinant; this isn't good, but we can live with it. 18.02 NOTES 6 Example 2.2 Find the inverse to A = Solution. We calculate that IAl = 2. Then the steps are (T means transpose): 1 0 -1 matrix A 1 0 1 -1 cofactor matrix T adj A inverse of A To get practice in matrix multiplication, check that A . A-l = I , or to avoid the fractions, check that A . adj (A) = 21. The same procedure works for calculating the inverse of a 2 x 2 matrix A. We do it for a general matrix, since it will save you time in differential equations if you can learn the resulting formula. matrix A cofactors T adj A inverse of A 2 2 Example 2.3 Find the inverses to: 3 2 1 2 0 Solution. a) Use the formula: IAI = 2, so A-' = 2 (-3 1) = 1 (4 $)0 . b) Follow the previous scheme: Both solutions should be checked by multiplying the answer by the respective A. Proof of formula (13) for the inverse matrix. We want to show A . A-l = I , or equivalently, A . adj A = IAJI; when this last is written out using (13) (remembering to transpose the matrix on the right there), it becomes To prove (14), it will be enough to look at two typical entries in the matrix on the right say the first two in the top row. According to the rule for multiplying the two matrices on the left, what we have to show is that M. MATRICES AND LINEAR ALGEBRA 7 These two equations are both evaluating determinants by Laplace expansions: the first equation (15) evaluates the determinant on the left below by the cofactors of the first row; the second equation (16) evaluates the determinant on the right below by the cofactors of the second row (notice that the cofactors of the second row don't care what's actually in the second row, since to calculate them you only need to know the other two rows). The two equations (15) and (16) now follow, since the determinant on the left is just JAl, while the determinant on the right is 0, since two of its rows are the same. The procedure we have given for calculating an inverse works for n x n matrices, but gets to be too cumbersome if n > 3, and other methods are used. The calculation of A-' for reasonable-sized A is a standard package in computer algebra programs and MatLab. Unfortunately, social scientists often want the inverses of very large matrices, and for this special techniques have had to be devised, which produce approximate but acceptable results. Exercises: Section 1G 3. Cramer's rule (some 18.02 classes omit this) The general square system and its solution may be written When this solution is written out and simplified, it becomes a rule for solving the system A x = b known as Cramer's rule. We illustrate with the 2 x 2 case first; the system is The solution is, according to (17), If we write out the answer using determinants, it becomes Crarner's rule: The formulas in the 3 x 3 case are similar, and may be expressed this way: Crarner's rule. If IAI # 0, the solution of A x = b is given by l Ail xi = 1A.I ' where lAil is the determinant obtained by replacing the i-th column of IA( by the column vector b. 8 18.02 NOTES Cramer's rule is particularly useful if one only wants one of the xi, as in the next example. Example 3.1. Solve for x, using Cramer's rule (19): Solution. We rewrite the system on the left below, then use Cramer's rule (19): I-: I: -7 Proof of (19). Since the solution to the system is x = A-'b, when we write it out explicitly, it becomes All A21 A31 A13 A23 A33 We show that this gives formula (19) for X I ; the arguments for the other xi go similarly. From the definition of matrix multiplication, we get from the above according to the Laplace expansion of the determinant by its first column. But this last equation is exactly Cramer's rule for finding x l . Cramer's rule is also valid for n x n systems; it is not normally used for systems larger than 3 x 3 however. You would use A-l, or systematic elimination of variables. Nonetheless, the formula (19) is important as a theoretical tool in proofs and derivations. Exercises: Section 1H 4. Theorems about homogeneous and inhomogeneous systems. On the basis of our work so far, we can formulate a few general results about square systems of linear equations. They are the theorems most frequently referred to in the applications. Definition. The linear system Ax = b is called homogeneous if b = 0; otherwise, it is called inhomogeneous. M. MATRICES AND LINEAR ALGEBRA Theorem 1. Let A be an n x n matrix. (20) (21) IAl IAl A x = b has the unique solution, x = A - l b . A x = 0 has only the trivial solution, x = 0. #0 #0 Notice that (21) is the special case of (20) where b = 0 . Often it is stated and used in the contrapositive form: A x = 0 has a non-zero solution (21') +- IAl = 0. (The contrapositive of a statement P +- Q is not-Q +- not-P; the two statements say the same thing.) Theorem 2. Let A be an n x n matrix. (22) (23) IAl = 0 IAl = 0 +- +- A X = 0 has non-trivial (i.e., non-zero) solutions. A x = b usually has no solutions, but has solutions for some b. In (23), we call the system consistent if it has solutions, inconsistent otherwise. This probably seems like a maze of similar-sounding and confusing theorems. Let's get another perspective on these ideas by seeing how they apply separately to homogeneous and inhomogeneous systems. Homogeneous systems: A x = 0 has non-trivial solutions w IAl = 0. Inhomogeneous systems: A x = b has the unique solution x = A-lb, if (A1 # 0. If IA( = 0, then A x = b usually has no solutions, but does have solutions for some b . The statements (20) and (21) are proved, since we have a formula for the solution, and it easy to see by multiplying A x = b by A-I that if x is a solution, it must be of the form = A-'b. We prove (22) just for the 3 x 3 case, by interpreting it geometrically. We will give a partial argument for (23), based on both algebra and geometry. Proof of (22). We represent the three rows of A by the row vectors a , b, c and we let x = (x,y,z); think of all these as origin vectors, i.e., place their tails at the origin. Then, considering the homogeneous system first, (24) Ax =0 is the same as the system a . x = 0, b . x = 0, c . x = 0. In other words, we are looking for a row vector x which is orthogonal to three given vectors, namely the three rows of A. By Section 1, we have IAl = a . b x c = volume of parallelepiped spanned by a , b , c. It follows that if IAl = 0, the parallelepiped has zero volume, and therefore the origin vectors a , b , c lie in a plane. Any non-zero vector x which is orthogonal to this plane will then be orthogonal to a,b, c, and therefore will be a solution to the system (24). This proves (22): if JAl = 0, then A x = 0 has a nontrivial solution. 18.02 NOTES 10 Partial proof of (23). We write the system as A x = d , where d is the column vector d = (dl, d2, d3)T . Writing this out as we did in (24), it becomes the system If IAl = 0, the three origin vectors a , b, c lie in a plane, which means we can write one of them, say c, as a linear combination of a and b: (26) c =ra + sb, r.s real numbers. Then if x is any vector, it follows that Now if x is also a solution to (25), we see from (25) and (27) that this shows that unless the components of d satisfy the relation (28), there cannot be a solution to (25); thus in general there are no solutions. If however, d does satisfy the relation (28), then the last equation in (25) is a consequence of the first two and can be discarded, and we get a system of two equations in three unknowns, which will in general have a non-zero solution, unless they represent two planes which are parallel. Singular matrices; computational difficulties. Because so much depends on whether IAl is zero or not, this property is given a name. We say the square matrix A is singular if IAl = 0, and nonsingular or invertible if IAl # 0. Indeed, we know that A-I exists if and only if IAl # 0, which explains the term "invertible"; the use of "singular" will be familiar to Sherlock Holmes fans: it is the 19th century version of "peculiar" or the late 20th century word "special". Even if A is nonsingular, the solution of Ax = b is likely to run into trouble if IAl z 0, or as one says, A is almost-singular. Namely, in the formulas given in Cramer's rule (19), the IAI occurs in the denominator, so that unless there is some sort of compensation for this in the numerator, the solutions are likely to be very sensitive to small changes in the coefficients of A, i.e., to the coefficients of the equations. Systems (of any kind) whose solutions behave this way are said to be ill-conditioned; it is difficult to solve such systems numerically and special methods must be used. To see the difficulty geometrically, think of a 2 x 2 system Ax = b as representing a pair of lines; the solution is the point in which they intersect. If IAl E 0, but its entries are not small, then its two rows must be vectors which are almost parallel (since they span a parallelogram of small area). The two lines are therefore almost parallel; their intersection point exists, but its position is highly sensitive to the exact positions of the two lines, i.e., to the values of the coefficients of the system of equations. Exercises: Section 1H