Factoring a polynomial can reveal invariant subspaces of an operator, but it is not enough merely to write down its factors. Why should the corresponding kernels add to the whole space, and why should their intersection be zero? The Chinese remainder theorem supplies explicit projection operators and answers both questions at once. Its module version also shows that the argument does not depend on choosing a basis.
Throughout, is a commutative ring with identity and is a unital -module.
Definition 1.5.1
(The submodule and comaximal ideals)
For an ideal , the submodule consists of finite sums of elements with and . Two ideals are comaximal when .
Remainders in a module
Theorem 1.5.2
(Chinese remainders for modules)
Let be pairwise comaximal ideals of , where . The map
is surjective and has kernel . Consequently it induces an isomorphism from to the displayed product.
The product has finitely many factors and can equally be written as a direct sum. The theorem prescribes compatible coordinates modulo different submodules; it does not assert that itself is their direct sum.
Proof
(Coefficients that separate the remainders)
Put , with empty product . Since for all , multiplying suitable identities with and shows . Choose such that . Then for . Given arbitrary representatives , the element maps to the desired tuple: its th coordinate differs from by an element of . This proves surjectivity.
The kernel is the intersection by definition. The product ideal clearly sends into every . Conversely let lie in the intersection. Since belongs to every , and , it sends into . The ideal equality follows by induction from for comaximal ideals, where for , , . Applying that ideal equality, and using , each lies in . Hence . The first isomorphism theorem finishes the proof.
Proof X-ray
What the coefficients do
Each behaves like one in the th remainder and zero in every other remainder. The choices need not be idempotent as elements of : they become orthogonal idempotents only after passing to the quotient by the product ideal, or when acting on a module annihilated by that ideal.
From factors to invariant kernels
Let be a field and linear; need not be finite-dimensional. Give its -module structure by . Suppose a nonzero polynomial annihilates and its factors are pairwise coprime. Define
Theorem 1.5.3
(Polynomial primary decomposition)
Under these hypotheses, . Each is -invariant. If is congruent to modulo and to modulo every other factor, then is the projection onto along the other summands.
Proof
(Projections from polynomial remainders)
The polynomial Chinese remainder theorem gives such and shows that , each , and each for are divisible by . Also is divisible by : it contains and all other pairwise coprime factors. Since , the corresponding operators satisfy , , and . Further, , so the image of lies in . Conversely, if , write ; then . Thus the image is exactly . The orthogonal projections sum to the identity, which gives both existence and uniqueness of the decomposition. Each kernel is invariant because commutes with .
When is the factorization into distinct irreducible polynomials, these summands are the primary components . The irreducibles need not be linear. For finite-dimensional , one may use a known annihilating polynomial such as the characteristic polynomial, by Cayley–Hamilton; a factorization of a polynomial that does not annihilate is insufficient.
Counterexample mode
Without an annihilating polynomial the components can vanish
Claim: every operator is the direct sum of its primary components . Take , infinite-dimensional, with multiplication by . For every nonzero polynomial and , whenever , because is an integral domain. So every component is zero, and their sum is . The repair is the stated hypothesis that some nonzero annihilates ; in finite dimension, Cayley–Hamilton supplies one.
How to find the projections
The existence proof is constructive. For two coprime polynomials and , the Euclidean algorithm produces polynomials with
Set and . The equation says and ; their factors say that and . Thus, if , every vector has the concrete decomposition
For example, . To see why the sum is direct, suppose lies in both kernels. Applying the Bézout identity gives . This separate intersection check is useful: writing a vector as a sum does not, by itself, prove uniqueness. The two operators and provide an algorithm for finding the summands rather than merely asserting their existence.
In the operator example below, let and . Since , the identity
shows exactly where the projection polynomials come from. The term containing is the projection onto , while the term containing projects onto . The factors are not interchangeable when labeling the target kernels.
A complete calculation
Worked example 1.5.4
(Separating a nilpotent block and an eigenvalue)
On let . Then , so . The factors and are coprime. Choose
The first is modulo and modulo ; the second has the reverse remainders. Thus and . Accordingly
The zero-primary component is two-dimensional even though is only one-dimensional. Replacing by the ordinary zero eigenspace would miss the vector and fail to span .
Common mistake
(Splitting a repeated factor)
The coprimality requirement cannot be omitted. For , the polynomial annihilates , but the two proposed kernels are the same line, . Their sum is not the plane and is not direct. Repeated factors must first be grouped into powers belonging to distinct irreducibles.
Worked example 1.5.5
(Reconstructing a residue from two pieces)
Take , and in . These ideals are comaximal because . The module quotient is naturally , and is naturally . For the desired pair , use , and , . The construction gives
Indeed has remainders and . If another residue of has both remainders, its difference from is divisible by both and , hence by , so it is the same element of . Here , making the remainder map an isomorphism on itself. In general the theorem only identifies the quotient by with the product.
Common mistake
(Expecting the remainder map itself to be an isomorphism)
The example with and can leave a false impression. There , so the map
is itself an isomorphism. That happens only because the kernel vanishes.
To see what the theorem says in general, we now take a module in which the kernel is not zero.
A module that is not the product
Let , and . Then and . We write for the coset of in , and for its coset in . These are cosets of different submodules and live in different quotient modules. For example, and are not comparable objects. The map is
Comaximality. Since , the element lies in . An ideal containing is the whole ring, so .
Surjectivity, concretely. The relation gives two useful integers.
- The integer lies in and is . Adding gives , which is still both. So and .
- The integer lies in and . So and .
Hence for any ,
This shows directly that is onto.
Worked example 1.5.6
(A preimage and all preimages)
Take the target . With and we get . Check: and , so .
Now suppose . Then , so and . Write . Then , and , so . Therefore . Conversely, every maps to .
The full preimage is therefore
The same computation shows , which agrees with the chapter's description . This kernel is not zero: although . So is not injective.
In fact no isomorphism exists, through or any other map. The target has exactly elements, while is infinite, so there is not even a bijection between them.
The chapter's quotient result still applies. induces a well-defined map
and this map is an isomorphism. It is the quotient that is a product, not .
Choosing representatives does not split
One might try to repair injectivity by restricting to the representatives . On that set is a bijection onto the target. Its inverse is a function
satisfying . Such a function is called a section of .
This does not make injective on , because for every . It also does not exhibit as a direct sum, because is not a homomorphism.
Counterexample mode
The section is not additive
Among , the integer that is and is . The integer that is and is . So
In the target, . If were additive we would need . But in we have .
The residues still agree: and . So and differ by , which is an element of the kernel.
The failure cannot be fixed by a cleverer choice of representatives. Suppose is any group homomorphism and put . Then
so and . Hence no homomorphism from the target back into can be a section of .
The structural distinction
Proposition 1.5.7
(Fibers of the remainder map)
For pairwise comaximal , the elements of with the same image under form exactly one coset of . Consequently, if , then is an isomorphism.
Proof
(Linearity turns fibers into cosets)
The kernel of is the set of with , that is, and .
Because is -linear, for we have
So the fiber over is precisely , a coset in , and one element of . It is not to be confused with the cosets or in the factors.
If , each fiber is the single point , so is injective. It is surjective by the chapter's result, hence it is a bijective module homomorphism, that is, an isomorphism.
Proof X-ray
Where comaximality enters
The argument has a single move: because is -linear, the question whether two elements have the same image becomes the question whether their difference lies in . Comaximality is used only through the theorem, which identifies that kernel with and supplies surjectivity; the fiber computation itself works for any linear map. Without comaximality the fibers are still cosets of , but that kernel can be larger than (for and on it is , not ), and surjectivity can fail.
In general, then, always holds. The map itself identifies with the product exactly when , as for . When , the map is not injective. In the example , the infinite module itself is not isomorphic to the finite product.
Worked example 1.5.8
(When ideals are not comaximal)
Take , and . The remainders and cannot come from one integer: the second remainder forces it to be odd. Thus the map is not onto. Its kernel is , whereas . Both conclusions of the theorem can fail without comaximality. This example does not say that they must fail for every non-comaximal pair or every module.
Exercises
Quick check
(Find all integers with and .)
Solution · Guided solution
Use the coefficients and found above. Put .
Check: and , so .
By the claim, the solution set is the coset . Since , this is , so the solutions are for . For example and .
Quick check
(Let , and in . Compute , and , and decide whether is an isomorphism.)
Solution · Guided solution
The ideals are comaximal because .
Computing . Multiplying by modulo gives . So , which has elements.
Computing . Multiplying the same by modulo gives . So , which has elements.
Computing . Since , we get .
Sizes of the factors. By Lagrange's theorem for the additive group, has elements and has elements. The target therefore has elements, the same as .
Conclusion. Because , the claim shows is injective. It is surjective by the chapter's theorem. Hence is an isomorphism, and here itself, not merely a quotient of it, is the product.
Practice and solutions
Quick check
(Show explicitly that by a remainder map, and find the preimage of .)
Solution · Answer
Map to . The ideals and are comaximal, so the map is an isomorphism. The number has the required remainders; every other solution differs from it by a multiple of .
Quick check
(Suppose on a vector space over a field of characteristic different from two. Find the projections onto the two kernels and prove that is diagonalizable, even if the space is infinite-dimensional.)
Solution · Answer
The coprime factors are and . The projections are and , respectively. Their images are the corresponding eigenspaces and their sum is . Choosing a basis for each eigenspace gives a basis of eigenvectors for their direct sum, including in infinite dimension.
Quick check
(Why does the preceding conclusion fail in characteristic two? Give a two-dimensional example.)
Solution · Answer
In characteristic two the factors coincide. Let on , where . Then , but is only one-dimensional. The sole possible eigenvalue is , so has no eigenbasis.
Quick check
(On let . Show that and find the two primary components and their projection operators. Why is this decomposition not a decomposition into real eigenspaces?)
Solution · Answer
Here , so sends a vector into the first two coordinates, which kills. The factors and are coprime. Since takes the value at , take and . Their operators send to and , respectively. Thus the components are and . The first plane has no nonzero real eigenvector: a real eigenvalue would satisfy . The irreducible quadratic factor therefore produces a two-dimensional real primary component rather than two real eigenspaces.
Quick check
(With the non-comaximal ideals of , take instead . Is the remainder map onto, and does its kernel equal ? Explain why your answer does not contradict the counterexample above.)
Solution · Answer
Multiplication by , , and is invertible on . Hence , both quotient modules are zero, and the remainder map to is onto with kernel . Comaximality is a sufficient condition for the theorem for every module; this calculation shows it is not necessary for these conclusions for a particular module.