Beruflich Dokumente
Kultur Dokumente
Contents
1 Eigenvalues and Eigenvectors
1.1 1.2 1.3 1.4 The Basic Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Some Slightly More Advanced Results About Eigenvalues Theory of Similarity and Diagonalization
1
1 4 5 8
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A x = 0,
for
an
nn
matrix and
and
each
n1
column
For reasons that will become more apparent soon, a more general version of this question which is also of interest is to solve the matrix equation problem corresponds to
A x = x,
where
= 0.) T :V V x
does
In the language of linear transformations, this says the following: given a linear transformation from a vector space
1.1
an
corresponding scalar
vector of
with
A x = x
is called an eigenvector of
A,
and the
A. 0
an eigenvector.
Important note: We do not consider the zero vector For a xed value of
the set
S whose elements are the eigenvectors x with A x = x, together with V . (This set S is called the eigenspace associated to the eigenvalue .)
S contains the zero vector. S is closed under addition, because if A x1 = x1 and A x2 = x2 , then A (x1 + x2 ) = (x1 + x2 ). [S3]: S is closed under scalar multiplication, because for any scalar , A (x) = (A x) = (x) = (x).
[S1]: [S2]:
It turns out that it is fairly straightforward to nd all of the eigenvalues: because
nn
we know precisely when there will be a nonzero vector is not invertible, or, in other words, when
t.
p(t)
of the matrix
A,
A. tn
Notation 1: Some authors instead dene the characteristic polynomial as the determinant of the matrix
than
tI A.
Notation 2: It is often customary, when referring to the eigenvalues of a matrix, to include an eigenvalue the appropriate number of extra times if the eigenvalue is a multiple root of the characteristic polynomial. Thus, for the characteristic polynomial
t2 (t 1)3 ,
= 0, 0, 1, 1, 1
if we
Remark: The characteristic polynomial may have non-real numbers as roots. Non-real eigenvalues are absolutely acceptable; the only wrinkle is that the eigenvectors for these eigenvalues will also necessarily contain non-real entries. (If
A has real number entries, then any non-real roots of the characteristic poly-
nomial will come in complex conjugate pairs. The eigenvectors for one root will be complex conjugates of the eigenvectors for the other root.)
This statement follows from the observation that the determinant of an upper-triangular matrix is the product of the diagonal entries, combined with the observation that if of times.)
A is upper-triangular,
then
tI A
is also upper-triangular. (If diagonal entries are repeated, the eigenvalues are repeated the same number
Example: The eigenvalues of 2, and 3.
1 0 0
i 3 0
3 8
are 1, 3, and
2 0 0
0 3 0
1 2 2
are 2,
A,
tI A
p(t).
of
p(t)
equal to zero and solve. The roots are precisely the eigenvalues
A. x form
satisfying
A x = x.
(I A) x = 0 via row-reduction.)
and the nonzero vectors in the space are the eigenvectors corresponding
. A= 1 0 0 1
.
Example: Find all eigenvalues, and a basis for each eigenspace, for the matrix
Step 1: We have
tI A =
t1 0
0 t1
, so
= 1, 1
(Alternatively, we could have used the fact that the matrix is upper-triangular.) Step 3: We want to nd the vectors with
a b
.
have this
is given by
1 0 A=
and
0 1 1 1
.
Example: Find all eigenvalues, and a basis for each eigenspace, for the matrix
1 0
Step 1: We have
tI A =
t1 0
1 t1
, so
= 1, 1 a+b b
(Alternatively, we could have used the fact that the matrix is upper-triangular.) Step 3: We want to nd the vectors with which means
. This requires
a b
= a 0
b = 0.
, and so
=1
is given by
1 0
1 0
1 1
1 0
0 1
polynomial and eigenvalues, but do not have the same eigenvectors. In fact, for for
= 1,
the eigenspace
1 0
1 1
1 0
0 1
is 2-dimensional.
Example: Find all eigenvalues, and a basis for each eigenspace, for the matrix
A=
2 3
2 1
tI A =
t2 3
2 t1
, so
= 1, 4 2a + 2b 3a + b
For
= 1
we want
a b
, so we need
a b
, which reduces
to
2 a = b. 3 =1
For
we want
2 3
2 1
a b
=4
a b
, so we need
a = b.
b b
, so a basis is given by
Example: Find all eigenvalues, and a basis for each eigenspace, for the matrix
0 A= 1 0 t 1 1 t
0 0 0 1 . 1 0 = t (t2 + 1).
Step 1: We have
t tI A = 1 0
0 t 1
0 1 , t
so
p(t) = det(tI A) = t = 0, i, i
.
a a 0 0 0 1 b = 0 b , so we need a c = 0 , so a = c and For = 0 we want c c b 0 0 1 a b = 0. So the vectors we want are those of the form 0 , so a basis is given by 0 . 1 a 0 0 0 a a 0 ia For = i we want 1 0 1 b = i b , so we need a c = ib , so a = 0 0 1 0 c c b ic 0 0 and b = ic. So the vectors we want are those of the form ic , so a basis is given by i . c 1 a a 0 ia 0 0 0 For = i we want 1 0 1 b = i b , so we need a c = ib , so 0 1 0 c c ic b 0 a = 0 and b = ic. So the vectors we want are those of the form ic , so a basis is given by c 0 i . 1
Example: Find all eigenvalues, and a basis for each eigenspace, for the matrix
1 A = 1 1
0 1 1 3 . 0 3 t1 0
0 t1 0
1 t1 3 , so p(t) = (t1) 0 t3
1 3 +(1) 1 t3
= 1, 2, 2
a+c a For = 1 we want so we need a + b + 3c = b , so a + 3c c 0 0 c = 0 and a = 0. So the vectors we want are those of the form b , so a basis is given by 1 . 0 0 1 0 1 a a a+c 2a For = 2 we want 1 1 3 b = 2 b , so we need a + b + 3c = 2b , so 1 0 3 c c a + 3 c 2c c a = c and b = 2c. So the vectors we want are those of the form 2c , so a basis is given by c 1 2 . 1
1.2
k.
appears as a root of the characteristic polynomial is called the , and the dimension of the eigenspace corresponding to is called the geometric
So what the theorem says is that the geometric multiplicity is at most the algebraic
(t 1)3 (t 3)2 ,
=1
is at most
=3
is at most 2-dimensional.
Proof: The statement that the eigenspace has dimension at least 1 is immediate, because (by assumption)
is a root of the characteristic polynomial and therefore has at least one nonzero eigenvector associated
to it.
k,
(I A) x = 0. If appears k times
I A B must have at most k rows of all zeroes. Otherwise, the matrix B (and hence I A too, although this requires a check) would have 0 as an eigenvalue more than k times, because B is in echelon form and therefore upper-triangular. But the number of rows of all zeroes in a square matrix is the same as the number of nonpivotal
as a root of the characteristic polynomial, then when we put the matrix into its reduced row-echelon form
B,
then
columns, which is the number of free variables, which is the dimension of the solution space.
So, putting all the statements together, we see that the dimension of the eigenspace is at most
k.
Theorem: If
A (a1 v1 + + an vn ) = A 0 = 0. Now since v1 , . . . , vn are eigenvectors this says a1 (1 v1 ) + + an (n vn ) = 0. But now if we scale the original equation by 1 and subtract (to eliminate v1 ), we obtain a2 (2 1 )v2 + a3 (3 1 )v3 + + an (n 1 )vn = 0. Since by assumption all of the eigenvalues 1 , 2 , . . . , n were dierent, this dependence is still nontrivial, since each of j 1 is nonzero, and at least one of a2 , , an is nonzero. But now we can repeat the process to eliminate each of v2 , v3 , . . . , vn1 in turn. Eventually we are left with the equation b vn = 0 for some nonzero b. But this is impossible, because it would say that vn = 0, contradicting our denition saying that the zero vector is not an eigenvector. So there cannot be a nontrivial dependence relation, meaning that v1 , . . . , vn are linearly independent.
Multiply both sides by the matrix
A:
this gives
Corollary: If
is an
nn
matrix with
distinct eigenvalues
v1 , v2 , . . . , vn
1 , 2 , . . . , n , and v1 , v2 , . . . , vn Rn . v1 , v2 , . . . , vn
they are a basis.
are (any)
This result follows from the previous theorem: it guarantees that so since they are vectors in the
n-dimensional A
vector space
Rn ,
is the determinant of
A.
p(t) = (t 1 ) (t 2 ) (t n ), we see that the constant term (1)n 1 2 2 . But the constant term is also just p(0), and since p(t) = det(tI A) we n have p(0) = det(A) = (1) det(A). Thus, setting the two expressions equal shows that the product of the eigenvalues equals the determinant of A. A
equals the trace of
A.
Note: The trace of a matrix is dened to be the sum of its diagonal entries. Proof: If we expand out the product
p(t) = (t 1 ) (t 2 ) (t n ) we see that the coecient of tn1 is equal to (1 + + n ). If we expand out the determinant det(tI A) to nd the coecient n1 of t , we can show (with a little bit of eort) that the coecient is the negative of the sum of the diagonal entries of A. Therefore, setting the two expressions equal shows that the sum of the eigenvalues equals the trace of A.
1.3
such that
n n matrices A and B are similar (or conjugate) if there exists B = P 1 AP . (We refer to P 1 AP as the conjugation of A by P .) A= 3 2 1 1
and
an invertible
nn
B = 2 1 3 2
1 1
2 1 3 2
P = =
P 1 =
2 1 P 1 AP = B .
3 2
1 1
2 1
3 2
2 3 1 2 1 2 1 1 A
, so that , so that
Q=
0 1
1 2
also has
B = Q1 AQ. Q
with
and
are
Q1 AQ = B . B = P 1 AP
Similar matrices have quite a few useful algebraic properties (which justify the name similar). If and
D=P
CP ,
B + D = P 1 AP + P 1 CP = P 1 (A + C )P . BD = P 1 AP P 1 CP = P 1 (AC )P .
exists if and only if
The product of the conjugates is the conjugate of the product: The inverse of the conjugate is the conjugate of the inverse:
A1
B 1
exists, and
=P
P.
det(P
P ) = det(A).
The conjugate has the same characteristic polynomial as the original matrix:
tI P P
A P ) = det(P
In particular, a matrix and its conjugate have the same eigenvalues (with the same multiplicities). Also, by using the fact that the trace is equal both to the sum of the diagonal elements and a coecient in the characteristic polynomial, we see that a matrix and its conjugate have the same trace.
If
x is an eigenvector of A with eigenvalue , then P 1 x is an eigenvector of B with A x = x then B (P 1 x) = P 1 A(P P 1 )x = P 1 A x = P 1 (x) = (P 1 x).
This is also true in reverse: if same eigenvalue). In particular, the eigenspaces for
eigenvalue
if
is an eigenvector of
then
P y
is an eigenvector of
(with the
A. B
that
One question we might have about similarity is: given a matrix similar to?
A,
is
A.
1 , , n ,
the simplest form we could plausibly hope for would be a diagonal matrix
1
.. .
A.
if it is similar to a diagonal matrix
with
A is diagonalizable D = P 1 AP . 2 3 6 7
D;
A=
P =
1 1
2 1
and
P 1 =
1 1
2 1 A
, then we have
P 1 AP =
4 0
0 1
= D.
then it is very easy to compute any power of
If we know that
D = P 1 AP ,
A:
Since Then
is diagonal,
Dk
k
k th
powers of
D. 2 1
D = (P
AP ) = P
(A )P , 6 7
so
A =P D P
Example: With
A = 2 1 =
2 3
as above, we have
Dk =
4k 0
0 1
, so that
Ak =
1 1
4k 0
0 1
1 1
2 4k 1 + 4k
2 2 4k 1 + 2 4k
k which are not positive integers. For 7 3 4 1 2 example, if k = 1 we get the matrix 1 , which is actually the inverse matrix A . And 3 4 2 1 0 2 2 6 2 if we set k = we get the matrix B = , whose square satises B = = A. 1 3 3 7 2
Observation: This formula also makes sense for values of
Theorem: An
nn
matrix
Proof: If
has
v1 , , vn
1 , , n ,
| P = v1 |
| |
| vn |
A.
| n vn . |
is invertible, and
AP = P D.
| If P = v1 |
| | | | | | | vn then AP = P D says Av1 Avn = 1 v1 n vn , which | | | | | | | (by comparing columns) says that Av1 = 1 v1 , . . . , Avn = n vn . Thus the columns v1 , , vn of P are eigenvectors, and (because P is invertible) they are linearly independent. n
distinct
Finally, the last statement in the proof follows because (as shown earlier) a matrix with eigenvalues has
Advanced Remark: As the theorem demonstrates, if we are trying to diagonalize a matrix, we can run into trouble if the matrix has repeated eigenvalues. However, we might still like to know what the simplest form a non-diagonalizable matrix is similar to.
The answer is given by what is called the Jordan Canonical Form (of a matrix): every matrix is similar
J1 J2
.. .
, Jn
where each
J1 , , Jn
of the form
J =
1
.. .
, 1
with
1 2 0
2 0
1 2
The existence and uniqueness of the Jordan Canonical Form can be proven using a careful analysis of
k.
(Regular
Roughly speaking, the idea is to use certain carefully-chosen generalized eigenvectors to ll in for the missing eigenvectors; doing this causes the appearance of the extra 1s appearing above the diagonal in the Jordan blocks.
p(x)
A,
then
p(A)
is the zero
(where in applying a polynomial to a matrix, we replace the constant term with that constant times
A=
2 3
2 1
, we have
det(tI A) =
t2 3
2 t1
= (t 1)(t 2) 6 = 10 9 6 7
t2 3t 4. 6 9 6 3
We can compute
A2 = 0 0
.
10 9
6 7
A2 3 A 4 I =
4 0
0 4
0 0
Proof (if
is diagonalizable): If
A is A.
D = P 1 AP
with
diagonal, and
p(x)
be
D are the eigenvalues 1 , , n of A, hence are roots of the characteristic p(1 ) = = p(n ) = 0. Then, because raising D to a power just raises all of its diagonal entries to that power, we can see 0 p(1 ) 1 .. .. .. that p(D ) = p = 0. = = . . . 0 p(n ) n Now by conjugating each term and adding the results, we see that 0 = p(D) = p(P 1 AP ) = P 1 [p(A)] P . So by conjugating back, we see that p(A) = P 0 P 1 = 0.
The diagonal entries of polynomial of
A.
So
A
of
D;
p(J ) = 0,
1.4
A = P 1 DP ),
Step 1: Find the characteristic polynomial and eigenvalues of Step 2: Find a basis for each eigenspace of Step 3a: Determine whether
A.
dimension (namely,
A.
the number of times the corresponding eigenvalue appears as a root of the characteristic polynomial) then the matrix is diagonalizable. Otherwise, the matrix is not diagonalizable.
D.
Example: For
A=
0 3
2 5
with
D = P 1 AP ,
Step 1: We have
tI A =
t 2 3 t 5 = 2, 3. 0 3
so
For
= 2 we need to solve
For
= 3 we need to solve
2 a a 2b 2a =2 , so = and thus a = b. 5 b b 3a + 5 b 2b b 1 form so a basis for the = 2 eigenspace is . b 1 2 0 2 a a 2b 3a =3 , so = and thus a = b. 3 5 b b 3a + 5 b 3b 3 2 2 b form so a basis for the = 3 eigenspace is . 3 3 b A
is diagonalizable, and
D=
.
2 0
0 3
. We
P = 0 3
1 1 2 5
2 3 1 1
To check: we have
P 1 =
3 1
2 1
, so
P 1 AP =
3 1
2 1
2 3
2 0
0 3
D.
D=
3 0
0 2
which diagonal matrix we want as long as we make sure to arrange the eigenvectors in the correct order.
1 1 0 Example: For A = 0 2 0 , determine whether there exists a diagonal matrix D and an invertible 0 2 1 1 matrix P with D = P AP , and if so, nd them. t1 1 0 t2 0 t2 0 so det(tI A) = (t1) Step 1: We have tI A = 0 = (t1)2 (t2). 2 t 1 0 2 t 1 The eigenvalues are therefore = 1, 1, 2.
Step 2:
0 a a ab a 0 b = b , so 2b = b and thus b = 0. 1 c c 2b + c c a 1 0 The eigenvectors are of the form 0 so a basis for the = 1 eigenspace is 0 , 0 . c 0 1 2a ab a a 1 1 0 For = 2 we need to solve 0 2 0 b = 2 b , so 2b = 2b and thus 2c 2b + c c c 0 2 1 b a = b and c = 2b. The eigenvectors are of the form b so a basis for the = 2 eigenspace is 2b 1 1 . 2 1 0 0 Step 3: Since the eigenspace for = 1 is 2-dimensional, the matrix A is diagonalizable, and D = 0 1 0 0 0 2 1 0 1 0 1 . We have three linearly independent eigenvectors, so we can take P = 0 0 1 2 1 1 0 1 1 0 1 1 0 1 0 1 To check: we have P 1 = 0 2 1 , so P 1 AP = 0 2 1 0 2 0 0 0 1 = 0 1 0 0 1 0 0 2 1 0 1 2 1 0 0 0 1 0 = D. 0 0 2 1 For = 1 we need to solve 0 0 1 2 2 1 1 1 Example: For A = 0 1 1 , determine whether there exists a diagonal matrix D and an invertible matrix 0 0 1 P with D = P 1 AP , and if so, nd them. t 1 1 1 t 1 1 so det(tI A) = (t1)3 since tI A is upper-triangular. Step 1: We have tI A = 0 0 0 t1 The eigenvalues are therefore = 1, 1, 1.
Step 2:
For
=1
we need to solve
1 0 0
a = b = 0.
=1
is
not diagonalizable .
Well, you're at the end of my handout. Hope it was helpful. Copyright notice: This material is copyright Evan Dummit, 2012. You may not reproduce or distribute this material without my express permission.