Why turn an operator into a module?
Two square matrices are similar when they describe the same linear operator in different bases. Deciding whether two given matrices are similar is hard to do by hand. Eigenvalues help, but they are not enough: as we will see, the matrices and have the same eigenvalues and the same characteristic polynomial, yet they are not similar. Some matrices over have no real eigenvalues at all.
This chapter replaces the search for a good basis with a structural question. A linear operator on lets polynomials act on vectors, so becomes a module over the polynomial ring . This ring is a principal ideal domain, and you already know the structure theorem for finitely generated modules over a PID. Applying that theorem gives a complete list of similarity invariants and a canonical matrix in each similarity class.
Standing conventions. Throughout, is a field, is a finite-dimensional -vector space with , and is -linear. We write , where is the matrix of in any basis. This polynomial is monic of degree . It does not depend on the basis, because
The determinant is multiplicative over the commutative ring , which justifies this computation.
We cite the structure theorem in the following form. Let be a finitely generated module over a PID . Then
where each is a nonzero nonunit and . The rank and the ideals are uniquely determined by ; this is the invariant factor form. In the elementary divisor form, the torsion part is with prime. The multiset of ideals is unique. In , every nonzero ideal has a unique monic generator, so we always normalize and to be monic. The invariant factors are then unique polynomials, not merely unique up to units.
The module
Evaluation at is the map given by . It is a ring homomorphism sending to . Define
The module axioms follow directly from the homomorphism property; for instance . We write for with this module structure. Constant polynomials act as ordinary scalars, so the -vector space structure is recovered from . A subspace is an -submodule exactly when it is -invariant, since closure under implies closure under every .
Finite generation. Any -basis of generates , because .
Torsion. The space has dimension . So the operators are linearly dependent. This gives a nonzero with , and then for every . Hence the annihilator
is a nonzero ideal. Its monic generator is the minimal polynomial . If , the annihilator is all of , so .
Since a nonzero kills all of , the module has no free summand. A copy of cannot be annihilated by a nonzero polynomial, because is an integral domain. So in the structure theorem.
Similarity is module isomorphism
Theorem 3.2.1
(Similarity and module isomorphism)
Let and be linear. An -linear map is -linear from to if and only if . Consequently, if and only if there is a linear isomorphism with .
In particular, for acting on , the matrices and are similar if and only if .
Proof
(Intertwining is module linearity)
Suppose is -linear. Then .
Conversely, suppose . By induction on we get . By linearity, for every , which says . Additivity is already given.
A module isomorphism is precisely a bijective module homomorphism, which proves the second sentence.
For matrices, is equivalent to . This says that is -linear, and is invertible.
So classifying matrices up to similarity is the same as classifying finitely generated torsion -modules up to isomorphism.
Companion matrices
Definition 3.2.2
(Companion matrix)
Let be monic with . Its companion matrix is the matrix with 's on the subdiagonal, last column , and zeros elsewhere:
Equivalently, for , and . Some books use the transpose of this matrix. Either convention works, but a single computation must use one convention consistently.
Lemma 3.2.3
(Cyclic modules and companion matrices)
Let be monic of degree .
- By the division algorithm, every class in has a unique representative of degree at most . So is an -basis. In this basis, multiplication by has matrix .
- Conversely, suppose is a -invariant subspace with a basis in which has matrix . Then as modules.
Proof
(Division algorithm and a cyclic generator)
(1) For we have . At the last basis vector, .
(2) The matrix condition says for . It also says
The map , , is -linear. It is surjective, since its image contains every , and its kernel contains . So it induces a surjection . Both spaces have dimension , so this surjection is an isomorphism.
Proposition 3.2.4
(Characteristic polynomial of a companion matrix)
For monic of degree , .
Proof
(Expansion along the first row)
We use induction on . For , .
Let and expand along its first row, which is .
- Deleting row and column leaves , where . By induction this term contributes .
- Deleting row and column leaves an upper triangular matrix with every diagonal entry equal to . Its determinant is . This term contributes .
Hence the determinant is .
Invariant factors and the rational canonical form
Theorem 3.2.5
(Rational canonical form)
There are unique monic nonconstant polynomials , called the invariant factors of , such that
There is a basis of in which has the block diagonal matrix . If has matrix in some basis, with monic, nonconstant and , then and for every . Two operators, or two matrices, are similar if and only if they have the same invariant factors.
Proof
(From the structure theorem to companion blocks)
By the description of above, is finitely generated with . The structure theorem gives the decomposition, and monic normalization makes the unique. They are nonconstant because they are nonunits. (A constant would contribute only a zero summand.)
Choose an isomorphism from the direct sum to . The images under of the bases of the summands form a basis of . By part (1) of the cyclic-module lemma and the similarity theorem, has the stated block matrix in this basis.
For the uniqueness of the block form, suppose has matrix in some basis. Each block's basis vectors span a submodule, and part (2) of the cyclic-module lemma identifies it with . So . The uniqueness part of the structure theorem then forces and .
The similarity criterion follows from the similarity theorem.
The matrix is called the rational canonical form. Note what is and is not canonical here. The matrix is unique once we fix the companion convention and order the blocks by divisibility. The basis is not unique, because it depends on the choice of .
Corollary 3.2.6
(Characteristic and minimal polynomials from invariant factors)
Suppose . Then
Proof
(Determinants of blocks and annihilators of summands)
The determinant of a block diagonal matrix is the product of the block determinants. The companion determinant formula then gives the formula for .
For , the annihilator of a direct sum is the intersection of the annihilators of the summands. We have . Since every divides , the intersection equals .
Several facts follow at once:
- , which is the Cayley–Hamilton theorem.
- .
- Every irreducible factor of divides some , and therefore divides . So and have the same irreducible factors, possibly with different exponents.
The dimension-zero boundary. If , then : the list of invariant factors is empty. In that case , since the determinant of the matrix is (an empty product). Also , since . The formula has no meaning when , so this case must be stated separately.
Worked example 3.2.7
(A cyclic operator over Q and an explicit companion basis)
Let
The matrix is triangular, so . The minimal polynomial divides and has the same irreducible factors, so is either or . We test the first option:
So . There is a single invariant factor, , and the rational canonical form is
To find a basis, take . Then , , and . Let have columns ; then . We can check directly that
So , and hence .
The choice of matters only in that must not be killed by or by . Many vectors work. Different choices give different matrices , but always the same .
Worked example 3.2.8
(A diagonalizable matrix with two invariant factors)
Let over . We find its invariant factors, its rational canonical form and a basis realizing it.
We have . The product is diagonal with entries , and , so it is zero. A polynomial dividing but missing either factor or cannot kill , because the eigenvectors and would survive. Hence .
By the invariant-factor corollary, the last invariant factor is , and the remaining invariant factors multiply to . Each of them is nonconstant, so there is exactly one, namely . This satisfies , as required. The rational canonical form is
For a basis, take for the first block. For the second block, take and . Then , which matches the last column of . So the basis realizes the rational form.
Splitting fields: primary decomposition and Jordan form
Definition 3.2.9
(Jordan block)
For and , the Jordan block is the matrix with on the diagonal, on the superdiagonal, and elsewhere. In other words, and for .
Theorem 3.2.10
(Jordan form)
Suppose splits into linear factors over . Then there is a basis of in which has matrix
The multiset of pairs is uniquely determined by . The order of the blocks and the basis are choices.
Proof
(Existence from elementary divisors)
The elementary divisor form gives , with each monic irreducible. By part (1) of the cyclic-module lemma, has block companion matrix in a suitable basis. The companion determinant formula then gives . Since splits and factorization in is unique, each has the form .
Now fix one summand . Let be the class of for . These representatives have degrees , so they form a basis. Multiplication by satisfies , where we set . So its matrix is .
Uniqueness follows from the block-counting theorem below.
Primary decomposition. Assume splits. Call the exponent of in the algebraic multiplicity , and define the generalized eigenspace . Then , with the sum over the distinct roots of , and .
To see this, work block by block in a Jordan basis.
- On with , the matrix is triangular with nonzero diagonal . So it is invertible, and so are all its powers.
- On , the matrix satisfies , and .
Therefore is the span of the basis vectors in the -blocks. Its dimension is the total size of those blocks, which is because . The subspaces are intrinsic, since they are defined without reference to any basis. The Jordan basis chosen inside each is not intrinsic.
Counting blocks from kernel dimensions
Theorem 3.2.11
(Counting Jordan blocks from kernel dimensions)
Assume splits. Fix and set for , so . For every integer , in any Jordan form of :
- the number of -blocks of size at least is ;
- the number of -blocks of size exactly is .
Proof
(Kernels block by block)
The kernel of an operator on a direct sum of invariant subspaces is the direct sum of the kernels on the summands. So is a sum of contributions, one from each block.
A block with contributes , because is invertible there.
On a block , let . Then , so when , and when . Hence , and the block contributes . Therefore
The difference equals if and otherwise. This proves the first formula. Subtracting the count of blocks of size at least from the count of blocks of size at least gives the second formula.
Proof X-ray
Why kernel dimensions see block sizes
The argument has one structural move and one counting move. The structural move is that kernels split along an invariant direct sum, so is computed block by block; blocks for other eigenvalues drop out because is invertible there. The counting move is that on one -block the kernel of gains exactly one basis vector per step until it fills the block, which gives . First differences then detect size at least , and second differences detect size exactly . The numbers are defined from without any Jordan basis, which is why the same argument proves uniqueness of the Jordan form. The splitting hypothesis is needed only so that a Jordan form exists to be counted.
The numbers are defined from alone. So the block counts are the same in every Jordan form of , which proves the uniqueness in the Jordan form theorem. The theorem also gives a dictionary between multiplicities and block data:
- The geometric multiplicity equals the number of -blocks.
- The algebraic multiplicity equals the sum of the -block sizes.
- The exponent of in equals the largest -block size. This holds because the minimal polynomial of is , and the annihilator of a direct sum is the intersection of the annihilators.
In general for every root of , and exactly when all -blocks have size .
Margin Note
Ranks before vectors
To find a Jordan form, compute the ranks of until the kernel dimension reaches . The block sizes are then fixed; hunt for chain vectors only when the question asks for .
Worked example 3.2.12
(A 4x4 matrix with a single eigenvalue)
Let
Here , so the characteristic polynomial splits.
- The matrix has two nonzero, independent rows, so and .
- Multiplying out, has a single nonzero entry, a in position . So .
- Finally , so .
The differences are . So there are two blocks of size at least , one of size at least , one of size at least , and none of size at least . The Jordan form is
From this we read off: geometric multiplicity , algebraic multiplicity , and . In particular, is not diagonalizable.
An explicit Jordan basis. We have , and , so is a chain for the size- block. Next, . The vector is an eigenvector outside the chain, and indeed . Let have columns . Then .
Other choices also work, for example as the top of the chain, or as the eigenvector for any .
The rational form. The elementary divisors are and , so and . The rational canonical form is
When the characteristic polynomial does not split
Counterexample mode
A real matrix with no real Jordan form
Claim: every real square matrix is similar over to a Jordan matrix. Let , so has no real root. If for a real invertible and a real Jordan matrix , then is upper triangular, so would split over , a contradiction. Hence the splitting hypothesis in the Jordan form theorem cannot be dropped. What survives is the rational canonical form: is already in that form, with the single invariant factor . Over the polynomial splits as , and is similar to . The field over which we work is part of the question.
Common mistakes
Common mistake
(Characteristic and minimal polynomials are not complete invariants)
The matrices and both have . Their minimal polynomials are and , so they are not similar. The matrices and both have and . They are distinguished by , which is for the first and for the second. Their invariant factors are also different: versus .
Common mistake
(Geometric multiplicity counts blocks, not size)
The geometric multiplicity counts blocks. The algebraic multiplicity counts total block size. Only when all blocks have size do the two agree. Not every matrix diagonalizes; the single-eigenvalue example above is one. Over a field where does not split, there may be no Jordan form at all.
Common mistake
(Invariant factors, elementary divisors and bases)
The invariant factors form a divisibility chain. The elementary divisors are prime powers, one for each Jordan block in the split case. The canonical forms are unique (up to block order in the Jordan case). The bases that produce them are not.
Exercises
Exercise 3.2.13
(Prove that is diagonalizable over if and only if is a product of distinct monic linear factors over .)
CoreProof
Solution(4)
- 1
Dispose of the zero space
If , the empty basis diagonalizes and is the empty product of distinct linear factors. Assume from now on that .
- 2
Diagonalizable implies a squarefree split minimal polynomial
Let be diagonalizable with distinct eigenvalues . The operator kills every vector of an eigenbasis, so divides . A monic divisor of a product of distinct monic linear factors is again such a product.
- 3
Conversely, the characteristic polynomial splits
Let with the distinct. Every irreducible factor of is then linear, so splits and the Jordan form theorem applies.
Because: By the invariant-factor corollary, and have the same irreducible factors.
- 4
Every Jordan block has size one
Each exponent in equals , so every block is and the Jordan form is diagonal.
Because: The exponent of in is the largest -block size.
∎
Exercise 3.2.14
(A complex matrix has and . Determine its Jordan form and its invariant factors.)
Core
Solution(3)
- 1
Block sizes for
The sizes add up to and the largest is , so they are and .
Because: The sizes add up to the algebraic multiplicity, and the largest size is the exponent in .
- 2
Block sizes for
The sizes add up to and the largest is , so they are and . The Jordan form is
- 3
Assemble the invariant factors
The elementary divisors are , , and . Multiply the largest power of each prime together, then the next largest:
Check: and .
Exercise 3.2.15
(Let be nilpotent on and . (a) If , find the Jordan form of . (b) Show that no nilpotent on has .)
Synthesis
Solution(3)
- 1
Reduce to the eigenvalue
Since for some , the minimal polynomial is a power of , so by the invariant-factor corollary . The Jordan form and block-counting theorems apply with .
- 2
Part (a): read off the block counts
Key step
The differences are : three blocks of size at least , two of size at least , one of size at least . So the Jordan form is , whose sizes add up to .
Because: counts the blocks of size at least .
- 3
Part (b): the counts cannot increase
The number of blocks of size at least can only decrease as grows. The data give differences , and , so no such operator exists.