Suppose 1 2 1 2 A= 2 4 2 5 . 1 2 1 1 (a) Give a basis for N (A):
We proceed by elimination to reduce A to upper-triangular form:
r2 − 2r1 1 2 1 2 r3 − r1 1 2 1 2 1 2 1 2 r3 +r1 A= 2 4 2 5 −→ 1 −→ 1 = U, 1 2 1 1 −1 0 which immediately tells us that A is rank 2, and that the 1st and 4th columns are the pivot columns. We will then solve equations by divid- ing the variables into pivot and free variables, x = [p1 , f1 , f2 , p2 ]. The nullspace will therefore be 4 − 2 = 2 dimensional, and our basis will need 2 vectors.
To find our usual basis for N (A), the special solutions, we will set the free variables to [1, 0] and [0, 1] and solve for the pivot variables, which leads to the upper-triangular systems:
1 2 p1 −2 −1 = or , 1 p2 0 0 where the left-hand-side is the upper-triangular matrix in the pivot columns of U and the right-hand-side is minus the free columns. By backsubsti- tution, we get p2 = 0 in both cases and p1 = −2 or − 1, respectively. Plugging these into x = [p1 , f1 , f2 , p2 ], we get our “special” basis for N (A): −2 −1 1 0 N (A) = span of 0 , 1 .
0 0
1 1 (b) For what value or values (if any) of α does Ax = 2α have any solu- α tion x?
To check whether a solution exists, we apply the same elimination steps
from A → U to this right-hand-side, and check if it is zero in the 3rd row (matching the row of zeros in U ), which is ensures that it is in C(A). Hence:
2 Problem 2 (24 points): Give a basis for the nullspace N (A) and a basis for the column space C(A) for each of the following matrices: 1 2 (a) The one-column matrix A = 3 .
4 This matrix is obviously rank 1 (full column rank), so N (A) = {~0} and the basis for N (A) is the empty set {}: the nullspace is zero-dimensional 1 2 so it needs no basis vectors. A basis for C(A) is just 3 , the first
4 column of A (which is also the pivot column).
(b) The one-row matrix A = 1 2 3 4 .
This matrix is also rank 1 (full row rank), with 1 pivot column and 3 free columns. We can read off the special solutions, so the 3-dimensional nullspace N (A) has the basis −2 −3 −4 1 0 0 0 , 1 , 0 .
0 0 1
More explicitly, the special solutions are of the form (p1 , f1 , f2 , f3 ), where we set the free variables to (1, 0, 0), (0, 1, 0), and (0, 0, 1) (the columns of I) and solve for p1 , but since this is one equation in one variable we can do it by inspection: p1 is just equal minus the free column.
Since it has full row rank, the column space C(A) is all of R1 , and is
spanned by the pivot column 1 .
Note that in 18.06 we sometimes gloss over the distinction between R
(scalars) and R1 (1-component column vectors) and R1×1 (1 × 1 matri- ces). If you think of A here as a “row vector” or “covector” that takes dot products with [1, 2, 3, 4], then the output is in R rather than R1 and you might say that a basis is the number 1 . I will accept that answer as well. 1 2 3 4 1 2 3 4 (c) The 100-row matrix A = . . . . in which every row is 1 2 3 4 . .. .. .. .. 1 2 3 4
3 This also has rank 1—after elimination, all the rows after the first will be zero. So N (A) will be 3-dimensional and C(A) will be 1-dimensional.
The first thing to realize is that we are doing the same operation as in part (b), but we are repeating the output 100 times. This doesn’t change the nullspace, since if the first row of the output is zero then all of the rows are zero. So the nullspace basis is the same as in part (b), i.e. N (A) is spanned by the special solutions −2 −3 −4 1 0 0 0 , 1 , 0 .
0 0 1
The column space C(A) is spanned by the pivot column—the first column, here—of A, which is simply 1 1 ∈ R100 , .. . 1
i.e. 100 rows of 1’s.
4 Problem 3 (25 points): 1 Suppose that we are solving Ax = 2 . In each of the parts below, a 3 complete solution x is proposed. For each possibility, say impossible if that could not be a complete solution to such an equation, or give the the size m × n and the rank of the matrix A if x is possible. 1 2 (a) ~x = 3
4 Impossible. A would need to be a 3 × 4 matrix, but such a matrix would have rank ≤ 3 and hence could not have unique solutions (could not be full column rank). 1 1 1 2 −1 0 (b) ~x = + α1 + α2 for all real numbers α1 , α2 ∈ R 3 5 0 4 17 1 Possible. Awould need to be a 3 × 4 matrix of rank 2, in order to have a 2d nullspace.
1 1 (c) ~x = +α for all real numbers α ∈ R 2 2 Impossible. For α = −1, this would give ~x = ~0, which could not be a solution with a nonzero right-hand side.
1 1 (d) ~x = +α for all real numbers α ∈ R 2 −1 Possible. A would need to be a 3 × 2 matrix of rank 1, in order to have a 1d nullspace.
1 1 1 (e) ~x = + α1 + α2 for all real numbers α1 , α2 ∈ R 2 −1 −1 Possible: 3 × 2 of rank 1. Since the second two vectors are the same, α2 is redundant with α1 and this is equivalent to the previous part with α = α1 + α2 .
Note: There was a typographical error in this problem: I had meant
to make the two vectors linearlyindependent, i.e. to ask something like 1 1 1 ~x = + α1 + α2 . In this case the solution would 2 −1 1 have been impossible: A would need to be a 3 × 2 matrix, but to have a 2d nullspace it would need to have rank 0, which means that no non-zero
5 right-hand-side could have a solution. Equivalently, the second two vec- tors form a basis for R2 if they are linearly independent, so there is some value of α1 and α2 that cancels the (1, 2) vector and gives ~x = ~0, which cannot be a solution with a nonzero right-hand side.
6 Problem 4 (25 points): Let 1 2 −1 −1 5 B= 1 1 , C= 2 −1 , b = −8 . 1 1 1 2 −4 Compute: (CB)−1 b. (Hint: Remember what I said in class about inverting matrices!)
Solution: As usual, we don’t want to compute matrix inverses explicitly, we
want to solve linear systems. In this case (CB)−1 b = B −1 C −1 | {z }b, y | {z } x
where y = C −1 b is computed by solving Cy = b using backsubstitution (since
C is upper-triangular), and then x = B −1 y is computed by solving Bx = y using forward-substitution (since B is lower-triangular). No Gaussian elimination is required! Proceeding, we have 2 −1 −1 y1 5 2y1 = 5 + y2 + y3 = −2 =⇒ y1 = −1 2 −1 y2 = −8 =⇒ 2y2 = −8 + y3 = −10 =⇒ y2 = −5 , 2 y3 −4 2y3 = −4 =⇒ y3 = −2 | {z } | {z } | {z } C y b
solve this inother ways. You 0 −2 −1 could multiply CB together to obtain CB = 1 1 −1 , and then labori- 2 2 2 1 1 3 2 4 8 ously invert this, e.g. with Gauss–Jordan, to obtain (CB)−1 = − 12 1 4 − 18 , 0 − 21 1 4
7 and finally multiply this by b. But this is a lot more work than doing a sin- gle backsolve followed by a single forward-solve, and is much more error-prone. In class, I repeatedly emphasized that you almost never need to compute matrix inverses explicitly, and if you do so then you are probably making a mistake. Inverses are useful for algebraic manipulations, but when it comes time to finally calculate something you should read them as “solving a linear system.” Another way of viewing this is that, for n × n matrices, back/forward solves take ∼ n2 operations, but both multiplying CB and inverting the matrix take ∼ n3 operations, so if I had given you a larger matrix then the penalty of doing it the slow way would have been even more dramatic.
