NOTES ON JORDAN CANONICAL FORM MATH 5316, FALL 2012 LANCE D. DRAGER 1. Polynomials and Linear Transformations Fix a vector space π over the complex numbers C. We’ll denote the dimension of π by π. Let π : π → π be a linear transformation. We’re safe in assuming π ΜΈ= 0. If we have a polynomial π(π§) ∈ C[π§], say π(π§) = ππ π§ π + ππ−1 π§ π−1 + · · · + π1 π§ + π0 , we can plug π into the polynomial in place of π§ to get a linear operator π(π ). We interpret the constant term in the polynomial as π0 π§ 0 , so when we plug in π we get π0 π 0 = π0 πΌ, where πΌ is the identity operator. Thus, π(π ) = ππ π π + ππ−1 π π−1 + · · · + π1 π + π0 πΌ. We will omit the πΌ in this expression if no confusion will result. Thus we write π − 3 for π − 3πΌ. Studying the operators π(π ) will allow us to analyze the structure of π . We begin by showing there is a polynomial so that π(π ) = 0. Lemma 1.1. If π : π → π is a linear operator, there is a non-zero polynomial π(π§) ∈ C[π§] so that π(π ) = 0 Proof. The operator π is in the vector space πΏ(π, π ) of linear operators on π . We know that πΏ(π, π ) is isomorphic to the space of π × π complex matrices, which has dimension π2 , so the dimension of πΏ(π, π ) is π2 . Consider the following vectors in πΏ(π, π ), 2 πΌ, π, π 2 , . . . , π π . This is a list of π2 + 1 vectors in a π2 dimensional space, so these vectors must be dependent. Thus, there are complex numbers 1 ππ (not all zero) so that 2 π0 πΌ + π1 π + π2 π 2 + · · · + ππ2 π π = 0. If we let π(π§) be the polynomial 2 π(π§) = π0 + π1 π§ + π2 π§ + · · · + ππ2 π§ π π, we get a nonzero polynomial so that π(π ) = 0. Version Time-stamp: ”2012-11-06 15:03:07 drager”. 1We’ll denote √−1 by i. 1 2 LANCE D. DRAGER We define βπ = {π(π§) ∈ C[π§] | π(π ) = 0}. This is (pretty obviously) an ideal in C[π§], see Exercise 1.3. Since all ideals in C[π§] are principal, we can find the monic generator π(π§) of βπ . This is the monic polynomial of least degree that annihilates π . The polynomial π(π§) is called the minimal polynomial of π . Similarly, if π£ ∈ π , we can look at π₯π£ = {π(π§) ∈ C[π§] | π(π )π£ = 0}. This is again an ideal, and it’s nonzero because βπ ⊆ π₯π£ , i.e., a polynomial that annihilates everything annihilates π£. The monic generator of this ideal is called the minimal polynomial of π£, and will be denoted by ππ£ (π§). Since the minimal polynomial π(π§) of π is contained in π₯π£ , ππ£ (π§) must divide π(π§). This fact is so handy, we’ll display it. Proposition 1.2. The minimal polynomial ππ£ (π§) of a vector π£ ∈ π divides the minimal polynomial π(π§) of π . Exercise 1.3. Show that βπ and π₯π£ are ideals in C[π§]. Exercise 1.4. Let π be a subspace of π and define π¦π = {π(π§) ∈ C[π§] | π(π )π = 0} = {π(π§) ∈ C[π§] | π(π )π€ = 0 for all π€ ∈ π } (1) Show that π¦π is a nonzero ideal. Denote the monic generator by ππ (π§) (2) Show that ππ (π§) divides π(π§). (3) Show that if π€ ∈ π , ππ€ (π§) divides ππ (π§). It will be useful to know how big the degree of ππ£ (π§) can be. Proposition 1.5. If π£ ∈ π , there is a polynomial π(π§) that annihilates π£ and has deg(π(π§)) ≤ π. Consequently, the degree of ππ£ (π§) must be less than or equal to π. Proof. We’ve already used the basic idea. Consider the vectors π£, π π£, π 2 π£, . . . , π π (π£). This list has π + 1 vectors in it, and they are in the π-dimensional space π , so they are linearly dependent. Thus, there are constants ππ , not all zero, so that π0 π£ + π1 π π£ + π2 π 2 π£ + · · · + ππ π π π£. If we define π(π§) = π0 + π1 π§ + π2 π§ 2 + · · · + ππ π§ π , we get a polynomial of degree ≤ π so that π(π )π£ = 0. Since π(π§) is a monic (non-constant) polynomial, we can factor it into linear factors as (1.1) π(π§) = (π§ − π1 )π1 (π§ − π2 )π2 · · · (π§ − πβ )πβ , where π1 , π2 , . . . , πβ are the distinct roots of π(π§). We next want to determine what these roots are. Theorem 1.6. The roots of π(π§) are exactly the eigenvalues of π . NOTES ON JORDAN CANONICAL FORM MATH 5316, FALL 2012 3 Proof. First, suppose that π is an eigenvalue of π . Then there is a nonzero vector π£ so that (π − π)π£ = 0. Thus, the minimal polynomial of π£ must be ππ£ (π§) = π§ − π (why?). Since ππ£ (π§) divides π(π§), π is a root of π(π§). Thus every eigenvalue is a root of π(π§). Next, we need to show that every root of π(π§) is an eigenvalue. Write π(π§) as in (1.1). We want to consider one of the roots. There is nothing special about how we labeled the roots, so we may well call our root π1 . Consider the polynomial π(π§) = (π§ − π1 )π1 −1 (π§ − π2 )π2 . . . (π§ − πβ )πβ , i.e., we’ve pulled out one factor of (π§ − π1 ), so π(π§) = (π§ − π1 )π(π§). Since π(π§) has degree less than the degree of π(π§), it is not divisible by π(π§). Thus, π(π ) ΜΈ= 0. Saying this operator is not zero means that there is a vector π£ ΜΈ= 0 to that π(π )π£ ΜΈ= 0. But then (π − π1 )[π(π )π£] = [(π − π1 )π(π )]π£ = π(π )π£ = 0π£ = 0. Thus, π1 is an eigenvalue of π , with eigenvector π(π )π£. We can now rewrite π(π§) as (1.2) π(π§) = (π§ − π1 )β1 (π§ − π2 )β2 . . . (π§ − πβ )ββ where π1 , π2 , . . . , πβ are the distinct eigenvalues of π . Remark 1.7. Remember the notation for the exponents in (1.2) since we will be referring to them. 2. Some Tools The following Proposition is standard linear algebra. The proof is included for completeness. Proposition 2.1. Let π be a vector space. Suppose that we have linear operators π1 , . . . , ππ that satisfy the following conditions. (1) Each ππ is a projection operator, i.e., ππ2 = ππ . (2) π1 + π2 + · · · + ππ = πΌ. (3) If π ΜΈ= π, ππ ππ = 0. Let ππ be the image of ππ . Then, π = π1 ⊕ π2 ⊕ · · · ⊕ π π and ππ coincides with the projection of π onto ππ defined by the direct sum decomposition. Proof. We first want to show that any π€ can be written as a sum of elements in the ππ ’s. This is easy, by Condition (2) we have π€ = π1 π€ + π2 π€ + · · · + ππ π€, and ππ π€ ∈ ππ by definition. Next we need to show that if (2.1) 0 = π€1 + π€2 + · · · + π€π , π€π ∈ ππ , then each of the components is zero. Since π€π ∈ ππ = im(ππ ), we can find a vector π’π so that π€π = ππ π’π . Thus, we have (2.2) 0 = π1 π’1 + π2 π’2 + · · · + ππ π’π . 4 LANCE D. DRAGER Fix an index π and apply ππ on the left of both sides of (2.2). We get 0 = ππ π1 π’1 + ππ π2 π’2 + · · · + ππ ππ−1 π’π−1 + ππ2 π’π + ππ ππ+1 π’π+1 + · · · + ππ ππ π’π . By Condition (3), the terms where the indices are different are zero, so we get 0 = ππ2 π’π . But ππ2 π’π = ππ π’π = π€π , by Condition (1), so π€π = 0. Since π was arbitrary, we conclude that all the components in (2.1) are zero. Finally, note that the projection onto ππ defined by the direct sum decomposition is to take π€ = π€1 + π€2 + · · · + π€π , (2.3) π€π ∈ ππ , to π€π . By the same computation as above, if we apply ππ to π€, the result is π€π . Thus, the projections coincide. For our next utility, we need a little algebra. Recall that if π1 (π§), π2 (π§), . . . , ππ (π§) are polynomials, the ideal (π1 (π§), π2 (π§), . . . , ππ (π§)) they generate is the set of all combinations π1 (π§)π1 (π§) + π2 (π§)π2 (π§) + · · · + ππ (π§)ππ (π§), π1 (π§), . . . , ππ (π§) ∈ C[π§]. This ideal is also equal to (π(π§)) for some polynomial π(π§). If the polynomials are not all zero, π(π§) can’t be zero and we can make a choice for π(π§) by choosing the monic one. We call π(π§) the greatest common divisor of the polynomials π1 (π§), π2 (π§), . . . , ππ (π§), written as gcd(π1 (π§), π2 (π§), . . . , ππ (π§)) . Since the gcd is in the ideal (π1 (π§), π2 (π§), . . . , ππ (π§)), we have gcd(π1 (π§), π2 (π§), . . . , ππ (π§)) = π1 (π§)π1 (π§) + π2 (π§)π2 (π§) + · · · + ππ (π§)ππ (π§), for some π1 (π§), . . . , ππ (π§) ∈ C[π§] If gcd(π1 (π§), π2 (π§), . . . , ππ (π§)) = 1 the polynomials are called relatively prime. Proposition 2.2. The common roots of π1 (π§), π2 (π§), . . . , ππ (π§) are exactly the roots of π(π§) = gcd(π1 (π§), π2 (π§), . . . , ππ (π§)). Consequently, the polynomials are relatively prime if and only if they have no common roots. Proof. Suppose that π is a root of π(π§). Since π(π§) divides everything in the ideal (π1 (π§), π2 (π§), . . . , ππ (π§)), we have ππ (π§) = ππ (π§)π(π§), for some polynomial ππ (π§). But then ππ (π) = ππ (π)π(π) = ππ (π)0 = 0. Thus, π is a root of each ππ (π§). For the converse, suppose π is a root of all of the ππ (π§)’s. We have π(π§) = π1 (π§)π1 (π§) + π2 (π§)π2 (π§) + · · · + ππ (π§)ππ (π§) for some ππ (π§)’s, so π(π) = π1 (π)π1 (π) + · · · + ππ (π)ππ (π) = π1 (π§)0 + · · · + ππ (π)0 = 0. If the polynomials π1 (π§), π2 (π§), . . . , ππ (π§) have no common roots, then π(π§) has no roots—so it must be a nonzero constant. The monic version is 1. Conversely, if our polynomials are relatively prime, the gcd is 1, which has no roots, so the polynomials have no common roots. NOTES ON JORDAN CANONICAL FORM MATH 5316, FALL 2012 5 3. Generalized Eigenspaces In this section we will define the generalized eigenspaces2and show that the whole space π is the direct sum of the generalized eigenspaces. To begin, let ππ be an eigenvalue of π . We say a vector π£ ΜΈ= 0 is a generalized eigenvector belonging to eigenvalue ππ if (π − ππ )π π£ = 0 for some positive integer π. Obviously, for any positive integer π, (π − ππ )π+π π£ = (π − ππ )π [(π − ππ )π π£] = (π − ππ )π 0 = 0, so (π − ππ )π π£ = 0 for π ≥ π. For the moment let π be the smallest positive integer so that (π − ππ )π π£ = 0. Then the minimal polynomial of π£ is ππ£ (π§) = (π§ − ππ )π (why?). By Proposition 1.5, we must have π ≤ π. Thus, if (π − ππ )π π£ = 0 for any power π, we must have (π − ππ )π π£ = 0. With this in mind, for each eigenvalue ππ , we define πΊ(ππ ) = {π£ ∈ π | (π − ππ )π π£ = 0} = ker((π − ππ )π ). The subspace πΊ(ππ ) is called the generalized eigenspace belonging to the eigenvalue ππ . Recall that πΈ(ππ ) = ker((π − ππ )) is the eigenspace belonging to ππ . Clearly πΈ(ππ ) ⊆ πΊ(ππ ), but in general they are not equal. Exercise 3.1. Consider the linear transformation C3 → C3 given by multiplication by the matrix β‘ β€ 0 1 0 π΄ = β£0 0 1β¦ . 0 0 0 The only eigenvalue is 0. Find πΈ(0) and πΊ(0). Exercise 3.2. Show that there are no “generalized eigenvalues”, i.e., if π ∈ C and there is a nonzero vector π£ and a positive integer π so that (π − π)π π£ = 0, then π is an eigenvalue of π . The following observation is useful. Proposition 3.3. If ππ is an eigenvalue, πΊ(ππ ) = ker((π − ππ )βπ ). Recall that βπ is the exponent of (π§ − ππ ) in the minimal polynomial π(π§), see (1.2) Proof. First suppose π£ ∈ πΊ(ππ ). If π£ = 0, there is nothing to prove. If π£ ΜΈ= 0 then (π − ππ )π π£ = 0, so the minimal polynomial of π£ must be ππ£ (π§) = (π§ − ππ )π for some positive integer π ≤ π. But ππ£ (π§) divides π(π§), so π ≤ βπ . Thus (π − ππ )βπ π£ = 0. The other direction is trivial. If π£ ∈ ker((π − ππ )βπ ), then (π − ππ )π π£ = 0, since π ≥ βπ . Thus, π£ ∈ πΊ(ππ ). Another useful observation is the following Proposition. 2Bad terminology, but we’re stuck with it. 6 LANCE D. DRAGER Proposition 3.4. π and π be distinct indices, so ππ ΜΈ= ππ . Then πΊ(ππ ) ∩ πΊ(ππ ) = {0}. Proof. Suppose that π£ ∈ πΊ(ππ ). Then (π − ππ )βπ π£ = 0. Thus, the minimal polynomial ππ£ (π§) of π£ must divide (π§ − ππ )βπ . This means that ππ£ (π§) must be ππ£ (π§) = (π§ − ππ )π , for some integer 0 ≤ π ≤ βπ (the zero vector would have minimal polynomial (π§ − ππ )0 = 1). On the other hand, π£ ∈ πΊ(ππ ), so (π − ππ )βπ π£ = 0. Thus, the polynomial (π§ − ππ )βπ must be divisible by ππ£ (π§) = (π§ − ππ )π . Sine ππ ΜΈ= ππ , the only way this is possible is to have π = 0, i.e., ππ£ (π§) = 1. But then 0 = ππ£ (π )π£ = 1π£ = π£, so π£ = 0. Next, we develop a little machinery about commuting operators. If πΏ and π are linear maps π → π , we say they commute if πΏπ = ππΏ. The following simple observations are left to the reader. Proposition 3.5. Let π , π, and π be linear operators π → π . Then, the following properties hold. (1) π commutes with itself. (2) If π commutes with π and π , then π commutes with the products ππ and π π. (3) If π and π commute, π π commutes with π π for any powers π and π (which can be negative if the operator is invertable). (4) If π commutes with π and π , then π commutes with πΌπ + π½π for any scalars πΌ and π½. (5) If π commutes with π , then π commutes with any polynomial π(π ) in π . We can use these facts to prove the following useful Proposition. Proposition 3.6. Let π be a linear operator π → π that commutes with π and let ππ be an eigenvalue of π . Then ππΊ(ππ ) ⊆ πΊ(ππ ) ππΈ(ππ ) ⊆ πΈ(ππ ) We say that the subspaces πΊ(ππ ) and πΈ(ππ ) are invariant under π. Proof. Suppose that π£ ∈ πΊ(ππ ), which means that (π − ππ )π π£ = 0. To test if ππ£ is in πΊ(ππ ), we need to see if (π − ππ )π [ππ£] = 0. But π and (π − ππ )π commute, so (π − ππ )π [ππ£] = [(π − ππ )π π]π£ = π[(π − ππ )π π£] = π0 = 0. Thus, ππ£ ∈ πΊ(ππ ) The corresponding result for the eigenspaces is left to the reader. Corollary 3.7. The eigenspace πΈ(ππ ) and the generalized eigenspace πΊ(ππ ) are invariant under any polynomial π(π ) in π . We now state the Big Theorem. Theorem 3.8 (Big Theorem). Let π1 , . . . , πβ be the distinct eigenvalues of π . Then π = πΊ(π1 ) ⊕ πΊ(π2 ) ⊕ · · · ⊕ πΊ(πβ ), in words, π is the direct sum of the generalized eigenspaces. NOTES ON JORDAN CANONICAL FORM MATH 5316, FALL 2012 7 Let’s discuss the proof, stating some important facts as Lemmas. Consider the polynomials (3.1) ππ (π§) = β ∏οΈ (π§ − ππ )βπ , π=1 πΜΈ=π in other words, we take the minimal polynomial and remove the factor (π§ − ππ )βπ corresponding to the eigenvalue ππ . β Lemma 3.9. The polynomials {ππ (π§)}π=1 are relatively prime. Proof of Lemma. It will suffice to show our polynomials have no common roots. The only possible roots are the eigenvalues π1 , π2 , . . . , πβ . But π1 is not a common root, because it is not a root of π1 (π§), π2 is not a root of π2 (π§), and so forth. Since the ππ (π§)’s are relatively prime, we have (3.2) 1 = π1 (π§)π1 (π§) + π2 (π§)π2 (π§) + · · · + πβ (π§), for some polynomials π1 (π§), . . . , πβ (π§). We’ll use the following notation ππ (π§) = ππ (π§)ππ (π§) ππ = ππ (π ) = ππ (π )ππ (π ). Plugging π in for π§ in (3.2) we have (3.3) πΌ = π1 (π ) + π2 (π ) + · · · + πβ (π ) = π1 + π2 + · · · + πβ . Lemma 3.10. For each π, im(ππ ) ⊆ πΊ(ππ ). Proof of Lemma. Let π£ be a vector in π . We want to show that ππ π£ ∈ πΊ(ππ ). By Proposition 3.3, it will suffice to show that (π − ππ )βπ ππ π£ = 0. (3.4) But (π§ − ππ )βπ is exactly the factor we removed from π(π§) to get ππ (π§). Thus, (π§ − ππ )βπ ππ (π§) = ππ (π§)(π§ − ππ )βπ ππ (π§) = ππ (π§)π(π§) and then (π − ππ )βπ ππ = ππ (π )π(π ) = 0, so (3.4) is certainly true. Lemma 3.11. If π ΜΈ= π, ππ πΊ(ππ ) = 0. Proof of Lemma. The factor (π§ − ππ )βπ appears in ππ (π§). Thus, ππ (π§) = π(π§)(π§ − ππ )βπ for some polynomial π(π§). Thus, ππ πΊ(ππ ) = ππ (π )πΊ(ππ ) = π(π )(π − ππ )βπ πΊ(ππ ) = 0, since (π − ππ )βπ kills πΊ(ππ ). Lemma 3.12. If π ΜΈ= π, ππ ππ = 0. Proof of Lemma. This follows from Lemma 3.10 and Lemma 3.11. Lemma 3.13. Each ππ is a projection operator, i.e., ππ2 = ππ . 8 LANCE D. DRAGER Proof of Lemma. We have πΌ = π1 + π2 + · · · + πβ . Multiply this by ππ on the left. This gives ππ = ππ π1 + ππ π2 + . . . ππ ππ−1 + ππ2 + ππ ππ+1 + . . . ππ πβ . All the terms where the indices are not equal are zero, so we wind up with ππ = ππ2 . We’ve now shown that the ππ ’s satisfy all the requirements of Proposition 2.1. If we let ππ = im(ππ ) ⊆ πΊ(ππ ), we have (3.5) π = π1 ⊕ π2 ⊕ · · · ⊕ πβ We will be done if we show that ππ = πΊ(ππ ). To do this, suppose that π£ ∈ πΊ(ππ ). We have, of course, π£ = π1 π£ + π2 π£ + · · · + πβ π£. Consider ππ π£ for π ΜΈ= π. On the one hand, ππ π£ ∈ ππ ⊆ πΊ(ππ ). On the other hand, ππ = ππ (π ) is a polynomial in π . By Corollary 3.7, πΊ(ππ ) is invariant under ππ , so ππ π£ ∈ πΊ(ππ ). But then ππ π£ ∈ πΊ(ππ ) ∩ πΊ(ππ ) = 0, using Proposition 3.4. Since ππ π£ = 0 for π ΜΈ= π, we have π£ = ππ π£ ∈ ππ , which completes the proof that ππ = πΊ(ππ ). This completes our proof of the Big Theorem, Theorem 3.8. 4. The Jordan Decomposition We begin with a discussion of nilpotent matrices. A linear transformation π : π → π is nilpotent if π π = 0 for some positive integer π. Of course, if π π = 0 then π π = 0 for any π > π. We call the smallest positive integer π such that π π = 0 the degree of nilpotency of π . Another way to characterize π is π π = 0 but π π−1 ΜΈ= 0. We want to show that if the dimension of π is π and π is nilpotent then π π = 0, i.e., π ≤ π. One way to see this is the following Proposition, which is useful in its own right. Proposition 4.1. Let π£ be a vector in π and let π : π → π be a linear transformation. Suppose there is a positive integer π such that π π π£ = 0, but π π−1 π£ ΜΈ= 0. Then the π vectors π£, ππ£, π 2 π£, . . . , π π−1 π£ are linearly independent. Proof. Suppose that we have a relation (4.1) π0 π£ + π1 ππ£ + π2 π 2 π£ + · · · + ππ−1 π π−1 π£ = 0. We need to show that all of the coefficients are zero. To do this, first multiply (4.1) on the left by π π−1 . This give (4.2) π0 π π−1 π£ + π1 π π π£ + · · · + ππ−1 π 2π−2 π£ = 0. Since π π π£ = 0 for π ≥ π, this reduces to just π0 π π−1 π£ = 0. Since π π−1 π£ ΜΈ= 0, we conclude that π0 = 0. Equation (4.1) now reduces to π1 ππ£ + π2 π 2 π£ + · · · + ππ−1 π π−1 π£ = 0. NOTES ON JORDAN CANONICAL FORM MATH 5316, FALL 2012 9 We now multiply this on the left by π π−2 , which gives us π1 π π−1 π£ + π1 π π π£ + · · · + ππ−1 π 2π−3 π£ = 0. Again, all the terms but the first are zero, so π1 π π−1 π£ = 0, from which we can conclude that π1 = 0. Continuing in this way, we conclude that all the coefficients are zero. Proposition 4.2. If π : π → π is nilpotent and the dimension of π is π, then π π = 0, i.e., the degree of nilpotency of π is less than or equal to π. Proof I. Suppose that π π−1 ΜΈ= but π π = 0. Since π π−1 ΜΈ= 0, there is a vector π£ such that π π−1 π£ ΜΈ= 0. But then the π vectors π£, π π£, π 2 π£, . . . , π π−1 π£ are linearly independent, so π ≤ dim(π ) = π. Proof II. Suppose that π π = 0 but π π−1 ΜΈ= 0. Then the polynomial π§ π annihilates π , but π§ π−1 does not. Thus, the minimal polynomial of π is π§ π . We know the degree of the minimal polynomial must be ≤ π. We continue to investigate our fixed linear transformation π : π → π , where π has dimension π. The goal of this section is to prove the following Theorem, which often suffices to solve a problem without going to the full Jordan Form. Theorem 4.3 (Jordan Decomposition). If π : π → π is a linear transformation, there are unique linear transformations π and π from π to π so that the following conditions hold. (JD1) π = π + π . (JD2) ππ + π π, i.e, π and π commute. (JD3) π is diagonalizable. (JD4) π is nilpotent. We’ll divide the rest of this section into the proof of existence and the proof of uniqueness. 4.1. Proof of Existence. As usual, let π1 , π2 , . . . , πβ be the distinct eigenvalues of π . From our Big Theorem, we have (4.3) π = πΊ(π1 ) ⊕ πΊ(π2 ) ⊕ · · · ⊕ πΊ(πβ ). Let π : π → π be the linear transformation that is given on πΊ(ππ ) by multiplication by ππ . Thus, if π£ ∈ π is decomposed as π£ = π£1 + π£2 + · · · + π£β , with respect to the direct sum decomposition (4.3), we have ππ£ = π1 π£1 + π2 π£2 + · · · + πβ π£β . Another way to say it is that (4.4) π = π1 π1 + π2 π2 + · · · + πβ πβ . Since the ππ ’s are polynomials in π , we see that π is a linear combination of polynomials in π , and so is a polynomial in π . Thus, π commutes with π , which is also easy to check from the definition of π. 10 LANCE D. DRAGER Exercise 4.4. Use the fact that the generalized eigenspaces are invariant under π to show that ππ = π π. We now define π = π − π. It’s clear that π commutes with both π and π . Indeed, π is a polynomial in π . We need to show that π is nilpotent. To do this, suppose that π£ ∈ πΊ(ππ ). We then have (π − π)π£ = π π£ − ππ£ = π π£ − ππ π£ = (π − ππ )π£. Since (π − π) commutes with (π − ππ ), we have (π −π)2 π£ = (π −π)[(π −π)π£] = (π −π)(π −ππ )π£ = (π −ππ )[(π −π)π£] = (π −ππ )2 π£. Continuing in this way, we get (π − π)π π£ = (π − ππ )π π£. Thus, π π π£ = (π − π)π π£ = (π − ππ )π π£ = 0, since π£ ∈ πΊ(ππ ). For an arbitrary π£ ∈ π , we have π£ = π£1 + π£2 + · · · + π£β , π£π ∈ πΊ(ππ ), and so π π π£ = π π π£1 + π π π£2 + · · · + π π π£β = 0 + 0 + · · · + 0 = 0. We’ve now constructed π and π satisfying the required four properties, so the proof of existence is complete. 4.2. Proof of Uniqueness. Denote the Jordan Decomposition we have constructed in the last subsection as π = πold +πold . Suppose that we have two transformations π and π so that π = π + π and (JD1) through (JD4) hold. We want to prove that π = πold and π = πold . Notice that π , π and π must all commute with each other. First, let’s determine the eigenvalues of π. Suppose that π is an eigenvalue of π and let π£ be an eigenvector for π. Then (π − π)π£ = π π£ − ππ£ = π π£ − ππ£ = (π − π)π£. Since (π − π) and π − π commute, we have (π − π)2 π£ = (π − π)[(π − π)π£] = (π − π)[(π − π)π£] = (π − π)[(π − π)π£] = (π − π)2 π£. Continuing this argument, we have (π − π)π π£ = (π − π)π π£ for any power π. But then (π − π)π π£ = (π − π)π π£ = π π π£ = 0, since π is nilpotent. From this we conclude, as in Exercise 3.2, that π is an eigenvalue of π . We can also conclude that π£, which started out as an eigenvector of π, is in the generalized eigenspace of π belonging to π. This gives us the following Lemma. Lemma 4.5. Let π be an eigenvalue of π. Then π is an eigenvalue of π and πΈπ (π) ⊆ πΊπ (π), i.e., the eigenspace of π is contained in the generalized eigenspace of π . NOTES ON JORDAN CANONICAL FORM MATH 5316, FALL 2012 11 Next, let π be an eigenvalue of π with eigenvector π£. Then (π − π )π£ = ππ£ − π π£ = ππ£ − ππ£ = (π − π)π£. Since (π − π ) and (π − π) commute, we can use the same procedure as above to show that (π − π )π π£ = (π − π)π π£ for any exponent π. But π − π = −π , so (π − π)π π£ = (π − π )π π£ = (−π )π π£ = (−1)π π π π£ = 0, since π π = 0. As before, this implies that π is an eigenvalue of π. We’ve now shown that the eigenvalues of π and π are exactly the same, and for each eigenvalue ππ we have πΈπ (ππ ) ⊆ πΊπ (ππ ). Of course we have π = πΊπ (π1 ) ⊕ πΊπ (π2 ) ⊕ · · · ⊕ πΊπ (πβ ). Since π is diagonalizable, we have π = πΈπ (π1 ) ⊕ πΈπ (π2 ) ⊕ · · · ⊕ πΈπ (πβ ). We can then cite the following Lemma. Lemma 4.6. Let π be a vector space and suppose that π1 , π2 , . . . , ππ , π1 , π2 . . . , ππ are subspaces of π so that (1) ππ ⊆ ππ for all π = 1, 2, . . . , π. (2) π = π1 ⊕ π2 ⊕ · · · ⊕ ππ . (3) π = π1 ⊕ π2 ⊕ · · · ⊕ ππ . Then ππ = ππ , π = 1, 2, . . . , π. Proof. Of course, any vector π£ can be written uniquely as π£ = π£1 + π£2 + · · · + π£π , where π£π ∈ ππ We want to prove ππ = ππ for all π. Suppose that π£ ∈ ππ . Then its decomposition as above is (4.5) π£ = 0 + 0 + · · · + 0 + π£ + 0 + · · · + 0, i.e., all the components are zero except for the one in ππ , which is π£. But we can also write π£ with respect to the direct sum of the ππ ’s, so (4.6) π£ = π€1 + π€2 + · · · + π€π−1 + π€π + π€π+1 + · · · + π€π , where π€π ∈ ππ . But ππ ⊆ ππ , so each π€π ∈ ππ . Thus, in (4.6), π£ is written as a sum of components where the πth component is in ππ . There is only one way to do this, namely (4.5). Thus, we have π€π = 0 for π ΜΈ= π and π£ = π€π ∈ ππ . This shows ππ ⊆ ππ , so the proof is complete. Applying this to the case at hand, we conclude that πΈπ (ππ ) = πΊπ (ππ ), 12 LANCE D. DRAGER in other words, π is given by multiplication by ππ on πΊπ (ππ ). This is exactly the definition of πold in our construction, so we conclude that π = πold . Then, of course, π = π − π = π − πold = πold , so the proof of uniqueness is complete. Department of Mathematics and Statistics, Texas Tech University, Lubbock, TX 79409-1042 E-mail address: lance.drager@ttu.edu