Math 20580
March 22, 2021
1. Elementary row operations
Here is what each elementary row operation does to the
determinant.
(1) Permute the rows by τ : multiply the determinant by
(−1)
τ.
(2) Multiply a row by c: multiply the determinant by c.
(3) Add a multiple of row j to row i: no change in the
determinant.
We will verify (3) later but first, find the determinant of
Row reduce:
1 2 −3 2 2 5 −1 7 5 12 −4 17 −1 −3 3 3 ∼ 1 2 −3 2 0 1 5 3 0 2 11 7 0 −1 0 5 ∼ 1 2 −3 2 0 1 5 3 0 0 1 1 0 0 5 8 ∼ 1 2 −3 2 0 1 5 3 0 0 1 1 0 0 0 3Since we only did type (3) elementary row operations, det(A) = 3
Following a different path to REF, let D = det(A).
2. Laplace expansion
The Laplace expansion of a determinant gives another way to calculate
which has theoretical consequences and can be a fast way to compute
if your matrix has a lot of 0’s. It comes about by asking for the matrix
of the linear transformation d
j: R
n→ R. From general theory there
exists a row vector with entries [C
1 j· · · C
n j] such that
det
Ah ~
Ai
j= [C
1 j· · · C
n j] ~
A
As usual C
i j= d
j(~
e
i) = det (Ah~
e
ii
j).
Appendix B checks that Ci j = (−1)i+jAi j so in general (ce) det(A) = [(−1)j+1|A1 j| · · · (−1)j+n|An j|A~j = n X k=1 Ck j xk j
If you work out the matrix for the row i linear transformation you will get (re) det(A) = Ri (−1)i+1|Ai 1| ... (−1)i+n|Ai n| = n X k=1 xi k Ci k The book shows you the checkerboard pattern of signs:
3. Determinants multiply
det(BA) = det(B) det(A) See Appendix C for a proof.
4. The adjugate of a matrix
The adjugate of a square matrix A is the matrix adj(A) whose (i j)th entry is Cj i, backwards to what you might expect. In other words adj(A) = [Ci j]T.
A adj(A) = det(A)1n×n adj(A)A = det(A)1n×n
See Appendix D for a proof. Notice that the adjugate is a pretty good substitute for the inverse because the adjugate always exists and is a ma-trix of polynomials in the xi j. If A is a matrix with det(A) 6= 0 then
A−1 = 1
det(A)adj(A).
5. Properties of the adjugate
It is easy to compute directly that adj(1n×n) = 1n×n and adj(0) = 0.
adj(cA) = cn−1adj(A) det adj(A) = det(A)n−1
adj(AB) = adj(B)adj(A) adj adj(A) = det(A)n−2A
These formulas can be guessed by pretending that adj(A) = A−1 and then adjusting by some determinants. Details may be found in Appendix E.
6. Cramer’s rule
Cramer’s rule gives a solution to the equation A~x = ~b if A is an invertible square matrix. It is very inefficient except for small n.
The formula comes from adj(A)A~x = adj(A)~b =⇒ det(A)~x = adj(A)~b. Hence det(A)xi is the ith row of adj(A)~b.
Appendix A. Type (3) elementary row operations So why does the formula for type (3) elementary row operations work?
Suppose we have A and we wish to add c times row j to row i. That is, if the rows of A are Rk, the new matrix C has rows Rk, k 6= i and row i is Ri+ cRj.
Look at the matrix B which has rows Rk, k 6= i and row i is cRj. Then one row of B is a multiple of another so det(B) = 0. By the row version of (P1) det(C) = det(A) + det(B).
Appendix B. The Laplace expansion
The point of Laplace expansion is that the Ci j can be worked out in terms of (n − 1) × (n − 1) determinants instead of the n × n ones which follow from the definition of the matrix associated to dj.
Here is one way to proceed. Let Ai j be the (i, j)th minor of A. That is the (n − 1) × (n − 1) matrix you get by deleting the ith row and jth column from A. Let Y be the matrix you get by replacing the jth column and the ith row by 0’s except for yi j = 1.
Looking at the matrix Ah~eiij we see that all terms in the Leibniz expansion
are 0 except for ones contain xi j. Look at xi k. Since xi j is present in any non-zero term, xi k is not present in any non-zero term unless k = j. Hence |Y| = |Ah~eiij| and Yi j = Ai j
To compute |Yi j| move the jth column to the far left. This introduces a
The new matrix looks like Yi j ~0 0 · · · 0 1 Hence Ci j = (−1)i+j Ai j ~0 0 · · · 0 1 = (−1)i+jAi j .
Then the Laplace expansion along the jth column is (ce) det(A) = [(−1)j+1|A1 j| · · · (−1)j+n|An j|A~j =
n
X
k=1
Ck j xk j where ~Aj is the jth column of A.
Either repeating the above argument using rows or playing with transposes yields the Laplace expansion along the ith row :
(re) det(A) = Ri (−1)i+1|Ai 1| ... (−1)i+n|Ai n| = n X k=1 xi k Ci k
Appendix C. Multiplicative properties of the determinant We discussed elementary matrices earlier in the course but here is a quick review.
Given any elementary row operation, one can apply it to the m × m identity matrix to get a matrix E which has the property that EA does the same elementary row operation on any m × n matrix A.
(1) If τ permutes the rows of 1 then multiplication by E(τ ) permutes the rows of A by τ : det E(τ ) = (−1)τ.
(2) If c is scalar and i is a row index, then multiplication by E(c, i) multiplies the ith row of A by c: det E(c, i) = c.
(3) If c is scalar and i and j are row indices, then multiplication by E(c, i, j) adds c times the jth row to the ith of A: det E(c, i, j) = 1.
Let E be any of the three kinds of m × m elementary matrices above. Then det(EA) = det(E) det(A)
Given any square matrix A there exists a sequence of elementary row oper-ations which reduce it to REF, R. Hence there is an E with det(EA) = det(E) det(A) = det(R). Now det(E) is not zero and det(R) is the product of the diagonal entries. All the diagonal entries of R are non-zero if and only if det(R) 6= 0 if and only if R is invertible if and only if A is invertible.
Hence the determinant determines if a square matrix is invertible. Moreover
det(BA) = det(B) det(A)
To see this observe that if det(B) 6= 0 B is a product of elementary matrices since there is such a product which reduces B to RREF which must be the identity matrix. Hence if det(B) 6= 0 we already known the equation holds. If det(B) = 0 then B is not invertible so there is a non-zero row vector R such that RB is the zero row vector. But then RBA is also the zero row vector and BA is not invertible so det(BA) = 0 and the equation holds in this case as well.
Appendix D. Basic property of the adjugate of a matrix To see the first formula, let A adj(A) = [si j] Then
si j = n
X
k=1
xi kCj k
Laplace expansion along the ith row is given by (re): det(A) =
n
X
k=1
xi k Ci k,
If i 6= j let Y have kth row the same as the kth row of A except that the ith row of Y is the jth row of A. Hence the entries of Y are yr s = xr s except yi k = xj k.
Let Cr s(A) be the indicated cofactor of A with Cr s(Y ) be the indicated
cofactor of Y. Since the rows of A and Y are the same except for row i, Ci k(A) = Ci k(Y ). det(Y) = n X k=1 yi k C(Y )i k = n X k=1
xj k C(A)i k = sj i. But Y has two equal rows so det(Y) = 0 and hence sj i = 0 and
A adj(A) = det(A)1n×n A similar proof using columns shows
adj(A) A = det(A)1n×n (1) If rank(A) = n, rank adj(A) = n.
(2) If rank(A) = n − 1 then rank adj(A) = 1.
(3) If rank(A) < n−1 then rank adj(A) = 0 which means adj(A) = 0n×n. (1): If rank(A) = n, A is invertible, so det(A) 6= 0 so det adj(A) 6= 0 and hence rank adj(A) = n.
(3): If rank(A) < n − 1, there are (n − 2) columns of A, k and j such that k and j are linear combinations of the remaining vectors. Hence if you remove one column, the remaining columns are still linearly dependent. Hence the columns of each Ai j are dependent and therefore det(Ai j) = 0; Cj = 0 and
(2): If rank(A) = n − 1, (n − 1) columns are linearly independent and one column, say j, is a unique linear combination of the others. Let Y be the n × (n − 1) matrix which is A with the jth column deleted. The columns of Y are linearly independent or, equivalently, rank(Y) = n − 1. This means that (n − 1) rows of Y are independent. But there are n rows so row must be a linear combination of the independent rows. Suppose row i is such a row. Then Ai j has linearly independent rows and hence det(Ai j) 6= 0 as is Ci j
and so rank adj(A) > 1.
Then adj(A)A = det(A)1n×n = 0n×n. There are (n − 1) linearly independent columns of A and each of them is a null vector for adj(A). By the Rank theorem, rank adj(A) + dim null space(adj(A) = n so rank adj(A) 6 1 and therefore rank adj(A) = 1
Appendix E. Properties of the adjugate It is easy to compute directly that adj(cA) = cn−1adj(A).
det adj(A) = det(A)n−1
It follows from adj(A) A = det(A)1n×n that det(A) det adj(A) = det(A)n. It is a theorem that cancellation holds with multi-nomials so our formula holds. Our formula implies A is invertible if and only if adj(A) is invertible. Moreover adj(A) = det(A)A−1. Hence the adjugate behaves a lot like the inverse except the adjugate always exists.
Start with AYadj(AY) = det(AY)1n×n =⇒ adj(A)AYadj(AY) =
det(AY)adj(A) =⇒ det(A)1n×nYadj(AY) = det(AY)adj(A) =⇒
det(A)adj(Y)Yadj(AY) = det(AY)adj(Y)adj(A) =⇒
det(A) det(Y)adj(AY) = det(AY)adj(Y)adj(A). Cancel det(AY) to get the formula.
adj adj(A) = det(A)n−2A
A adj(A) = det(A)1n×n =⇒ adj A adj(A) = det(A)n−11n×n =⇒ adj adj(A)adj(A) = det(A)n−11n×n =⇒
adj adj(A)adj(A)A = det(A)n−11n×nA =⇒ adj adj(A) det(A) = det(A)n−1A. Cancel a copy of det(A) to get our formula.
adj(A)T = adj(AT)
A adj(A) = det(A)1n×n =⇒ A adj(A)T = det(A)1n×n =⇒ adj(A)TAT = det(A)1n×n =⇒ adj(A)TATadj(AT) =
det(A)1n×nadj(AT) =⇒ adj(A)
T
det(AT)1n×n = det(A)adj(AT) =⇒
det(A) adj(A)T = det(A)adj(AT). Cancel a copy of det(A) to get our formula.
Appendix F. Cramer’s Rule
The formula comes from adj(A)A~x = adj(A)~b =⇒ det(A)~x = adj(A)~b. Hence det(A)xi is the ith row of adj(A)~b.
det(A)xi =
n
X
k=1
det Ai(~b) =
n
X
k=1
Ck i Ai(~b)bk
but because we have removed the ith column in forming Ck i Ai(~b),
Ck i Ai(~b) = Ck i A) so det(A)xi = n X k=1 Ck ibk = n X k=1 Ck i Ai(~b))bk = det Ai(~b)
Cramer’s Rule: If A~x = ~b has a solution, then det(A)xi = det Ai(~b)
If det(A) 6= 0 Cramer’s rule computes each xi. But it does say more. If
det(A) = 0 and if A~x = ~b has a solution, then each det Ai(~b) = 0.