Fundamentals of Linear Algebra and Geometry

Lecture 12: Changing bases

We have seen that a linear map f : V → W between vector spaces of finite type is determined
by its matrix once we have fixed bases in V and W . In the previous lecture, we have focused on
maps between standard spaces Rn with the canonical choice of bases. It is natural, and quite
useful, to consider the more general case, which is non much more complicated.
In fact, the only thing that one should keep in mid is that the entries of the matrices are
coordinates with respect to the given bases.
So, for instance, if M = M (f ; B, C) is the matrix of f in bases B of V and C of W , then
the columns of M give the coordinates in C of the generators of im f . Any solution of the
homogeneous system M x = 0 gives the coordinates in B of a vector of ker f .
The essential formula for changing bases is composition formula (1) in Lecture 10. Let B and B 0
be two bases of V and C and C 0 be two bases of W . the the matrices M (f ; B, C) and M (f ; B 0 , C 0 )
are related by the formula:
M (f ; B 0 , C 0 ) = M (idW ; C, C 0 )M (f ; B, C)M (idV ; B 0 , B).
It may be helpful to visualize this formula by means of the following diagram:
f
(V, B) −−−−→ (W, C)
x 

idV 
id
y W
f
(V, B 0 ) −−−−→ (W, C 0 )
which gives two paths to go from V with basis B 0 to W with basis C 0 : either by the bottom
arrow, corresponding to the matrix M (f ; B 0 , C 0 ); or going up to V with basis B, then by f to W
with basis C and finally going down by the change of basis from C to C 0 . This triple composition
correspond to the triple multiplication in the formula.
Let us emphasize a special case to be investigated later. Suppose that V = W . A linear map from
a space to itself is called an endomorphism. For endomorphisms we will often want to have a
matrix with respect to the same basis in the source and target, i.e. a matrix of the M (f ; B, B).
We will in thse case simplify the notation and write M (f ; B). Then if B 0 is another basis,
the relation between M (f ; B) and M (f ; B 0 ) is M (f ; B 0 ) = M (idV ; B, B 0 )M (f ; B)M (idV ; B 0 , B).
We have already noticed in Lecture 10 that M (idV ; B, B 0 ) = M (idV ; B 0 , B)−1 . So, the relation
between the matrices of an endomorphism in two different bases is of the form
M (f ; B 0 ) = H −1 M (f ; B)H
where H = M (idV ; B 0 , B).
Definition Two n × n matrices A and B are similar if there exists an invertible n × n matrix
H such that B = H −1 AH.
So the matrices of an endomorphism in different bases are similar. Conversely, returning to the
point of view of lecture 11, we can say that B is similar to A if B is the matrix of fA : Rn → Rn
in the basis given by the columns of H.
We will examine the similarity relation in quite some detail.
These topics are detailed in § 4 and 5 of Chapter 2. Take the time to read these carefully. Then
proceed with the exercises below. Post your solutions, questions etc. in the thread Lecture 12
in the Exercises Forum.
Solutions will be published at the end of the day.
Exercise 1. Let f : R3 → R4 be the map f (x, y, z, t) = (x + y, 2x + 3y + z, x + z, 2x + z). Find,
if possible, a basis B of R3 such that
 
2 0 1
7 0 3
M (f ; B, E4 ) = 
3 2 2 .


4 3 3

Is the such a basis B unique?

Exercise 2. In M (2 × 2, R), consider the subspaces

     
a b a b
V1 = | a + b + c = 0; c − d = 0 ; V2 = | a + 2b + c = 0; c − d = 0 .
c d c d

Consider then the subspaces of R3 :

      
1 1  x 
W1 =<  2  ,  1  >; W2 =  y  | x − 2y = 0 .
0 1 z
 

a) Is there a linear map g : M (2 × 2, R) → R3 such that g(V1 ) = W1 and g(V2 ) = W2 ? Is it

unique?

b) Let g be a map as above. Choose bases for M (2 × 2, R) and R3 and write the matrix of g
with respect to these bases.

c) Write the matrix of g with respect to the canonical bases.

d) Find bases for ker g and im g.

 
2
e) Determine g −1  1 .
3

Solutions

Solution to Exercise 1. The matrix of f wrt the canonical bases is obtained by putting in
coulums the vectors f (1, 0, 0), f (0, 1, 0) and f (0, 0, 1). Plugging these in we get
 
1 1 0
2 3 1
A = M (f ; E3 , E4 ) = 1 0 1 .


2 0 1

The problem is now to find a 3 × 3 matrix H = M (idR3 ; B, E3 ) such that AH = M (f ; B, E4 ).

This is possible if and only if the columns of M (idR3 ; B, E3 ) generate the same subspace as those
of A. That requirs some (not too difficult) computations but it would not be enough because
we are requested to actually compute H. This can be done as follows: the first (resp. second,
third)n column of H is a solution of the system Ax = first (resp. second, third)n column of
M (idR3 ; B, E3 ). In order not to have to repeat the computation, let us (try to) solve the system
Ax = t (a, b, c, d). Notice that this system will not be solvable for arbitrary a, b, c, d: that would
mean that f : R3 → R4 is surjective, which is impossible by the dimensions’ formula. But there
is hope we can solve it for the columns of M (idR3 ; B, E3 ).
Reducing the augmented matrix to row-echelon form:
     
1 1 0 a 1 1 0 a 1 1 0 a
 2 3 1 b   0 1 1 b − 2a   0 1 1 b − 2a 
     .
 1 0 1 c   0 −1 1 c − a   0 0 2 b + c − 3a 
2 0 1 d 0 −2 1 d − 2a 0 0 3 2b + d − 6a
 
1 1 0 a
 0
 1 1 b − 2a 

 0 0 2 b + c − 3a 
−1
0 0 3 2 (3a − b + 3c − 2d)
Therefore the system is solvable if and only if 3a − b + 3c − 2d = 0. Luckily, that is the case for
the columns of M (idR3 ; B, E3 ). Notice that as a byproduct, we get that rk A = 3, so the solution
is unique. Substituting for (a, b, c, d) the columns of M (idR3 ; B, E3 ) we get

f −1 (1, 2, 1, 2) = (1, 1, 2); f −1 (0, 0, 2, 3) = (1, −1, 1); f −1 (1, 3, 2, 3) = (1, 0, 1).

Hence B = {(1, 1, 2), (1, −1, 1)(1, 0, 1)}.

Solution to Exercise 2. Let’s first determine bases for the subspaces:

     
a b 1 0 0 1
V1 = =< , >;
−a − b −a − b −1 −1 −1 −1
     
a b 1 0 0 1
V2 = =< , >.
−a − 2b −a − 2b −1 −1 −2 −2
From this, we see that dim V1 = dim V2 = 2. Furthermore, in view of the common equation
c = d, V1 + V2 6= M (2 × 2, R). On the other hand, V1 6= V2 , as for instance the second generator
of V1 is not in V2 . So 2 dim(V1 + V2 ) 4, and the only possibility is then dim(V1 + V2 ) = 3.
From the Grassmann formula, dim(V1 ∩ V2 ) = 1. Actually, the first generator of V1 is in V2 , so
 
1 0
V1 ∩ V2 =< >
−1 −1

The generators of W1 are non-proportional, so they are a basis and dim W1 = 2. The space W2
is given as the solutions of a homogeneous equation (system of rank 1), so dim W2 = 3 − 1 = 2.
The first generator of W1 is in W2 while the second isn’t. So W1 6= W2 , hence W1 + W2 = R3
and  
1
W1 ∩ W2 =<  2  > .
0
A linear map g : M (2×2, R) → R3 is uniquely determined by its values on a basis of M (2×2, R).
The conditions g(V1 ) = W1 and g(V2 ) = W2 force
 
  1
1 0
g (V1 ∩ V2 ) = W1 ∩ W2 i.e. g =α  2 .
−1 −1
0

The inclusion is clear, but must have equality, otherwise g would map the first generator of V1
to zero and then we can’t have equality g(V1 ) = W1 , because dim g(V1 ) would be strictly less
than dim V1 = dim W1 = 2. This means α 6= 0 in the formula above. We may pick α = 1 (this
is arbitrary, so our eventual g will not be unique) and further put
   
  1   0
0 1 0 1
g =  1 ; g =  0 
−1 −1 −2 −2
1 1
the latter being an element of W2 not proportional to the generator of W1 ∩ W2 . Since the three
matrices generating V1 + V2 are independent, there exists a linear map g : M (2 × 2, R) → R3
satisfying these 3 conditions. It is not unique, as the three matrices are not a basis of M (2×2, R).
Let’s pick any matrix not in V1 + V2 : it suffices to take one with c 6= d, for instance E22 . We can
map it anywhere in R3 . The simplest option is to map it ton zero. Now there exists a unique
linear map g such that:
       
  1   1   0   0
1 0 0 1 0 1 0 0
g =  1 ; g
 =  1 ; g
 =  0 ; g
 =  0 .
−1 −1 −1 −1 −2 −2 0 1
0 1 1 0
       
1 0 0 1 0 1 0 0
In the basis B = , , , of M (2 × 2, R) and
−1 −1 −1 −1 −2 −2 0 1
the canonical basis in R3 , we have
 
1 1 0 0
M (g; B, E3 ) =  2 1 0 0  .
0 1 1 0
To write the matrix in the canonical bases, we need to compute M (idM (2×2,R) ; E, B). It is the
inverse of the matrix whose columns are the coordinates of the matrices in B with respect to
the canonical basis of M (2 × 2, R), which are just their entries:
 
1 0 0 0
 0 1 1 0 
M (idM (2×2,R) ; B, E) = 
 −1 −1 −2 0  .


−1 −1 −2 1
We omit the tedious computation of the inverse (done by the method seen in Lecture 9). It is:
 
1 0 0 0
 1 2 1 0 
M (idM (2×2,R) ; E, B) = M (idM (2×2,R) ; B, E)−1 = 
 −1 −1 −1 0  .


0 0 −1 1
Hence the matrix of g in the canonical basis is M (g, E, E3 ) = M (g, B, E3 )M (idM (2×2,R) ; E, B):
 
  1 0 0 0  
1 1 0 0  2 2 1 0
1 2 1 0 
M (g, E, E3 ) =  2 1 0 0    −1 −1 −1 0  =
  3 2 1 0 .
0 1 1 0 0 1 0 0
0 0 −1 1

By construction, we know that im g = W1 + W2 = R3 , so we may take the canonical basis. By

the dimension formula, dim ker g = dim M (2 × 2, R) − dim im g = 4 − 3 = 1. We have decide to
map the matrix E22 to zero, so we conclude ker g =< E22 >.
Finally, to compute the inverse image of t (2, 1, 3) we have to solve the system whose augmented
matrix is
     
2 2 1 0 2 2 2 1 0 2 2 2 1 0 2
 3 2 1 0 1   0 −1 − 1 0 −2   0 1 0 0 3 
2
0 1 0 0 3 0 1 0 0 3 0 −1 − 12 0 −2
   
2 2 1 0 2 2    
 0 1 0 −1  −1 3 0 0
0 3  =⇒ g 1  = +< >.
−2 0 0 1
0 0 − 12 0 1 3