Unique Factorization Domains


We have already been familiar with the fundamental theorem of arithmetic every since our elementary school, that for every positive integer can be decomposed into product of postitive prime power form , and the decomposition is unique.

When it comes to the case of decomposing a polynomial in over , for a reduction of polynomial , it might also be decomposed into , although by the Gauss's lemma these two decompositions are fundamentally the same (and only the former one may be taken over ).

When it comes to (terms of degree higher than are absobred by either the term or the term ), for integer , there's both and , so the decomposition is seemly not unique here.

In this text, we will study when and why would a ring element in an integral domain be factorized uniquely, so that we will be able to make distinguishment of integral domains capable of unique factorization from those incapable of.

Unique factorization domains

In order to define the "factorization", let's have a closer look at the example of decomposition in (which is unique by fundamental theorem of arithmetic) and the one in and . The natural numbers are not closed under subtraction, so we prefer to study the whole of integers , while in , there's , although we perceive indifferently. In , there's , although we still perceive to be the same. So intuitively, in an integral domain , when talking about factorization, as we're always able to take out and cancel the units, we would like to make the elements such that equivalent, and are said to be associate.

Also, we borrow the concept of irreducibility of polynomials here. When talking about irreducibility of polynomials, we only care about how it can be decomposed into non-constant polynomials of lower degree, a constant polynomial multiplies the one associate with it trivially. In this way, for any non-unit non-zero element , if it can be written as product of non-unit non-zero elements that , then it is said to be reducible. Conversely, for an non-unit non-zero element that cannot be written as product of non-unit non-zero elements, then it's said to be irreducible. The units and zero are neither reducible nor irreducible.

The problem remains as, for every non-unit non-zero element in integral domain , it's either irreducible or can be written as product of other non-unit non-zero elements. We decompose these elements in the product recursively, until it all of the elements in the product are irreducible. And given two such factorization of :

Where are all irreducible. If there're and exists , such that and are associate pairwisely for , then these two factorization are equivalent. And for non-unit non-zero element , among all its factorizations, there's just a single equivalence class, then is said to be uniquely factorized in . And if every non-unit non-zero elements in can be uniquely factorized, is said to be a unique factorization domain, or UFD abbreviatedly.

Clarification of the introductory cases

Since we have now made the definition of "unique factorization" clear, we can revisit the examples provided in the introduction of this text.

The factorization of is unique: if is reducible, it can only be decomposed into two polynomial of degree ; it must be divisible by both and due to its roots in ; for any irreducible polynomial in potentially other factorization, it has root other than , so the other factor of the factorization must take the root both of the root , while the other factor of the factorization is of degree or to take roots, which is a contradiction. So the factorization is unique, and can be uniquely factorized in .

Now consider the example of that . First let's show that for any complex number, the the norm function defined by has a homomorphic property that , since there's:

Given that is a subring of , we can trim the domain of norm function to without affecting its homomorphic property, with the codomain trimmed to simultaneously. In short, we get the trimmed norm function defined by , with .

Any for any unit such that , the only possibility is , which can only be taken at , yielding to be the only units. There's also no zero divisor in , since if there were any , by it's only possible that or , which means it's either or . And is thus an integral domain.

For the irreducibility, assume there's some such that , then there must be either or . The other solution besides to is , which is associate with , and there's no solution to . For , there must be either or , the other solution besides to is , which is associate with , and there's no solution to . For , there must be either or . The only other solution besides to is , and it's associate with , and there's no solution to . The can also be tested in the same way. So none of the is reducible.

Finally, given there're two factorization of with no possible permutation to make their irreducible elements in the product associate pairwisely, cannot be uniquely factorized in , which is indeed not a UFD.

Principal ideals of the elements

When it comes to integral domain , a good thing is that is a commutative ring with unity, so the principal ideals are in the form of . And on the other hand, we could write that , although such notation would include both of the associativity and reducibility, given that we can add constraint on the factor if necessary, it won't hurt.

For two elements such that , given that , there're and thus .

A non-zero non-unit element is said to be prime once its principal ideal is prime ideal, that if then there's either or , which is equivalent to saying if then or , just like the Euclid's lemma but this is for the integral domain's case.

Every prime element must be irreducible, otherwise given that , there's but , which is a contradiction to 's being prime. Conversely an irreducible element is not necessarily prime, consider the example in , all of the are irreducible, but none of them is prime.

Essence of unique factorization

Given that we've talked about the prime elements, and there's irreducible element that is not prime in general, we would like to show it's necessary for all irreducible elements to be prime in a UFD.

Assume there's some irreducible element that is not prime in a UFD , so that there's some such that but . In this case, there's a irreducible factorization of that has a irreducible factor , and another irreducible factorization . Given that is associate with neither nor , the two factorization are not equivalent, thus cannot be uniquely factorized in , and is not UFD, which is a contradiction. In this way we've shown it's necessary for all irreducible elements to be prime in UFD.

Let's consider an integral domain such that all irreducible elements are prime, is it enough for us to derive the unique factorization? Consider the process that we take an element for factorization, first we find some non-unit non-zero elements such that , then we could like to dive into both and and decompose them in a depth-first traversal manner, until we've reach the level that we are required to decompose an irreducible element. But what if the factorization could last for forever, while in mathematics we don't care whether there's a factorization algorithm that will eventually terminate, but the factorization's lasting forever means the length of the irreducible factorization will be of infinite length. Back to the definition of UFD, we've made guarantee on unique factorization on only those of finite length, and the appearance of such decomposition of infinite length are out of scope and break our promise on the element's properties. So we must require the decomposition to be of finite length.

So now we have an integral domain such that all irreducible elements are prime, and the factorization of any non-unit non-zero element is of finite length. Is it enough now? Let's test it, first assume it's not a UFD by for element there're two non-equivalent factorizations:

Without losing generality, let's assume it's the case that .

Since is irreducible and is prime under such context, and there's , one of the , , ..., must hold, and without losing generality, let's assume it is the case that , so and must be associate, and let it be . Given that is an integral domain, there's no zero divisor, so we can safely cancel out from both sides, leaving that:

Let's repeat the process of cancelation exhaustively, until there're:

Assume , by associating all terms before into , we have , and thus is a unit, which is a contradiction to 's being an irreducible element. So it's only , and all these and are associate pairwisely. The two factorizations must be equivalent and thus can be uniquely factorized.

For now, we have reached the point to claim that if in an integral domain, all irreducible elements are prime and all non-unit non-zero elements have factorization series of finite length, then the integral domain is a UFD.

Principal ideal domains

Principal ideal domain, or abbreviatedly PID is an integral domain where all ideals are principal ideal. We have already encountered instances of such kind of integral domains: and .

The thing special about PIDs is that they fulfils both criteria for UFD. And we will prove why does it so in this section.

Irreducibility equals maximality in PID

In a PID , consider the principal ideal of an irreducible element . Assume there were an ideal such that , then there's . Since , cannot be a unit, and since , and are not associate. So it's only possible for to be a non-zero non-unit element that factorizes , however is irreducible, which is a contradiction. In this way, we simply shows that the principal ideals of irreducible elements are maximal.

Conversely, let there be a maximal ideal , assume there's some non-unit non-zero element such that , then we have , and thus is not maximal, which is a contradiction. So the maximal ideal can only be generated by an irreducible element.

For every irreducible element, its principal ideal is maximal, and is thereby prime. In this way, we've shown that every irreducible elements in a PID is prime, fulfiling our first criterion for UFD.

Ascending chain condition of principal ideals in PID

Let's think about the process of decomposition again, that . By setting as root, each time we decompose into product of two elements, like , we create a binary tree, where leaf nodes are primes. When it comes to PID, given that all ideals are principal ideals and the correspondence between elements and their principal ideals: the root node is the principal ideal of the element we want to decompose; every branch node has two children by , and we have ; every leaf node is a maximal or prime ideal. A good thing about the tree is that associate elements generate the same principal ideal. And in this way, when there are any two distinct binary tree of factorization, we could gather their leaf nodes, and they are just a permutation of another, otherwise there'll irreducible element that is not prime, which is a contradiction.

More generally, an ascending chain in is a chain of principal ideals in the form of . If we are able to prove that every ascending chain has finite length, then for any binary tree of factorization, let the longest path inside be of finite length , then there're at most nodes and between and leaves, and the tree is therefore finite. In this way, will fulfil the second criterion for UFD.

To prove, let's consider the summation of elements in an infinite ascending chain that . While such kind of stuff looks intimidating at the first glance, we would like to have a look at its set theoretic and topological properties first. A question we would like to think about carefully is, given a monotonically increasing sequence of , that and , what's the union of closed intervals ? Is it or ?

To show it's the latter case, think of any , since , by expressing it in language we have , by replacing with we will get the corresponding such that . But for the right endpoint , given that , there's no closed interval to include this point, so it will not be included in the final interval. In this way, we have and . Therefore, it's . More radically, each set gathers the elements that fulfils certain condition relative to , and there're and .

Look back to our case, to show that is an ideal of , consider:

  1. .
  2. , assume it's , then , so .
  3. , .

And since is an ideal of , it must be principal, and let it be . Of course there's , so . But if the ascending chain is infinite, there must be a successor , and by set inclusion there'll be , which is a contradiction.

In this way, we've shown there's no infinite ascending chain in a PID. Given that a PID fulfils both criteria for UFD, we've proved that PIDs are UFDs.

Euclidean domains

Besides all ideal's being principal, another common point for and is both of them have division algorithm defined on it. For specified dividend and divisor , the division algorithm yields quotient and remainder such that , and for some "metrics" , e.g. and , there's . The metric some how ensures that Euclidean algorithm based on the division will eventually terminate at , finding some for input and .

And we would like to define the Euclidean domain , which is an integral domain equipped with a function called the Euclidean valuation, and must fulfil:

  1. For any , we have .
  2. For any , always exists , such that with or .

Given that we can define valuation for , and valuation for , both and are Euclidean domains.

Inequality properties of Euclidean valuations

We would like to derive some basic but useful properties of Euclidean valuations before digging into deeper.

The first one is the valuation get its minimal value in . Given that , must be minimal among all elements in . Meantime, given that , all must also be the same, and minimal by .

The second one is for any , and , there's . To show, we divide by , and there's , with either or . If , since there's , and no zero divisor in , we can safely cancel out on both sides, leaving and to be unit then, which is the impossible case. When , given that there's while neither nor is zero, we have , and thus .

The third one is the valuation get its minimal value only in . Based on the previous property, we set , and there's . By combining with the first property that , we finally have .

More specially, the valuation is said to be multiplicative whenever there's . Given that , it's only possible for to be or . When , there's some non-unit non-zero such that , so it's only possible for to be in this case. Both the valuation for and valuation for are multiplicative.

Euclidean domains are PIDs

We would like to show Euclidean domains are PIDs. Since we've already done something similar on and , and the proof is just some easy abstraction for them.

The zero ring and the whole ring can be seen as being generated by or . For some non-trivial ideal , take any non-zero and let , we'll check whether there's , if so, we are done; otherwise take any , and let , and check whether . We do such kind of thing iteratively, and we could show such process will create a finite monotonic ascending chain of principal ideal with , in which serves as a reliable stop signal, and must be among them.

To show the existence of ascending chain of principal ideal, we prove by mathematical induction that starting from , for the principal ideal obtained at step , , yielding principal ideal on step with .

Since , the Euclidean algorithm yields non-zero value such that . And in detail, we divide by when , or divide by when . Whatever case yields . Besides the common divisor , the Euclidean algorithm also yield the Bezout's identity that . Utilizing the Bezout's identity, for any , , so there's , and thus . In this way, the result yielded by Euclidean algorithm is divisible by any common divisor of and , and is thus maximal.

On the other hand, for any , there's . Conversely, for any , there's . And thus we have . In such way, by letting , we have with .

In this way, we've shown all Euclidean domains are PIDs, and are therefore UFDs.

Application in algebraic integers

While Euclidean domains are fundamentally PIDs, but they provides convenient language for proving a given integral domain is a PID.

Consider the Gaussian integers where . Since is the subring of , we can make a trimmed norm function defined by with . With this map, we can simply show the units of are , and there's no zero divisor since .

We would also like to show that can be used as candidate for a multiplicative Euclidean valuation. In the first place, by , it fulfils the first criterion for a Euclidean valuation.

To show it also fulfils the second criterion for a Euclidean valuation, we need to define its division first: by dividing by , we find quotient and remainder such that . The way of doing such division is described in a geometrical way. First, given that all points of are in , and is a 2-dimensional vector space of , we can choose an orthogonal basis , , and represent as . Then, we find some integers such that , and is trapped in the square pinpointed by , aka , , , . Among these four vertices, we choose the closest vertex to , and denote it as . Since in the square of edges' length , the distance of any point inside (including edges and vertices) to its closest vertex is no more than , we have:

In this way, we've shown is truly a Euclidean valuation for , and is a Euclidean domain and thereby a UFD.

I guess you can't be more confused than now: we've shown not to be a UFD, but also shown the Gaussian integers to be a UFD. While and look exactly the same, and have homomorphic map defined in the same way, what makes diverge from the ?

Obviously, given that norm function trimmed to also fulfil and , there's for non-zero , so it fulfils the first criterion, and we need to check whether it also fulfils the second criterion. When dividing by , we adapt the same way of choosing orthogonal basis , and represent as . We pinpoint the "cage" with to trap , but this time it's in rectangle shape rather than square, given that . If we choose the closest vertex to , in the rectangle of edges' length and , the distance of any point inside (including edges and vertices) to its closest vertex is no more than , which is greater than . So can't be used as the Euclidean valuation for . And given that not a UFD by counterexample , there's no other possible Euclidean valuation and it's impossible to be a Euclidean domain.

The integral domain which is subring of is said to be norm-Euclidean when the trimmed norm function can be used as an Euclidean valuation, which in turns imply that is an Euclidean domain. However, not being norm-Euclidean does not necessarily mean it can't be Euclidean. For example, it can be proved that is Euclidean but not norm-Euclidean, although it's too complex and we will not prove it here.

Another classical example of norm-Euclidean domain is the Eisenstein integers , where is the cubic primitive root. In this case, the norm function is trimmed to defined by . When dividing by , we choose the basis , which tessellates the complex plane with rhombus "cages". Again, we choose the most proximate vertex among to . In a rhombus of edge length , the maximal distance to the nearest vertices is , which is reached at the mass center of the two equilateral triangles sticking together back-to-back to form the rhombus. So we have:

And the Eisenstein integers form norm-Euclidean domain.

Polynomial rings over UFDs

We've discussed about properties of polynomials with integer coefficients , deriving some useful conclusions like Gauss's lemma, Eisenstein's criterion, and so on. For a polynomial ring over a UFD , given that divisibility are defined, and irreducible elements are primes, it sheds light on generalizing these conclusions in to .

More importantly, we would like to show polynomial ring over a UFD is also a UFD which is generally not a PID: like , it's over which is not a field, and the ideal is not principal: the common divisors of its elements is , however is not inside (while it is not a problem for since there's and is inside); the polynomial ring over UFD with multiple indeterminates, whose ideal is known not to be principal, can also be shown to be UFD, by recursively breaking and descending it through , proving every intermediate polynomial rings to be UFDs in a last-in-first-out manner.

Gauss's lemma for polynomial rings over UFDs

For polynomial ring over UFD , polynomial is said to be primitive when there's no prime element , so that . And for any polynomial that is not primitive, we can iteratively take common prime factors (repetition is allowed) of coefficients out, so that there's where is called the content of and is primitive, called the primitive part of . The process should halt in finite step otherwise the coefficients of can be written as infinite product of prime elements, which is a contradiction to 's being UFD.

A primitive polynomial can only be factorized into primitive polynomial: let be reducible by and without losing generality, is not primitive by . Then given that , is not primitive, which is a contradiction. Conversely, the product of two primitive polynomials is also primitive, otherwise assume there were , so that there's or , and assume at least the former case is true. Since cannot divide every coefficient of , let be the coefficient such that while , and be the coefficient such that while . Consider the coefficient :

Which is a contradiction. So the product of primitive polynomial can only be primitive.

Since is also an integral domain, we can also extend and embed into its field of fractions , so that we have , but we can also draw a conclusion that a primitive polynomial is reducible over iff it's reducible over . First we would like to show that every polynomial is convertible into where and is primitive: let's take all denominators out so that there's , noted that there's since is embedded into injectively by ; then we into its content and parts so that there's ; finally we have and we are done. Then for some polynomial that is reducible over by with , let's do such process for so that there's , let where there's no common prime element such that . If both are units, we are done. Otherwise let , if is unit while is not, for some prime , by observing that , we have so is not primitive; if is not unit, since , there must be , for any prime element such that , it's only possible that , leading to 's not being primitive. So the only possibility is that both are units. In this way, is reducible over by with . Conversely, a factorization of with is also a factorization in by . So the reducibility of primitive polynomial in is just the same as the one in .

For those who are familiar with Gauss's lemma for polynomials with integer coefficients, these proofs are nothing more than simple migrations into polynomials over UFDs.

Eisenstein's criterion for polynomial rings over UFDs

Let's make a quick work of it. The Eisenstein's criterion for over UFD states that a primitive polynomial is irreducible when there's a prime such that . Assume is reducible by , since there's , and while implies it's either or . Assume it's the former case, so there's a coefficient such that while , but there's , which is a contradiction. So polynomial must be irreducible when it fulfils the Eisenstein's criterion.

Polynomial rings over UFDs are UFDs

For any polynomial , we will first factorize its content and primitive part, and then furtherly factorize content and primitive part correspondingly. Assume there're two distinct factorizations:

Where are the factorization of content and must be irreducible elements over ; and are the factorization of primitive part and must be irreducible polynomials over .

Such two factorizations over are also factorizations over . We've proved that all ideals of a polynomial ring over field are principal, so is a PID and thus a UFD. The elements and are elements of and thus units in , while irreducible polynomials are irreducible elements, so there must be while exists such that and are associate by .

For two primitive polynomials that are associate in , obviously there be and , while it still remains unknown whether are non-units but coprime. Assume it's currently such case, however, since are polynomials in , for each coefficients there's . For any prime , it's the case that , and therefore is not primitive; for any prime , it's the case that , and therefore is not primitive. Both cases are impossible, so it's only possible that are units, and .

In this way, it's actually the case that . Their products and are also associate in (as is said, and are elements of and thus units in ), and by Gauss's lemma they are also primitive, so there's , and therefore . Since is UFD, it's only possible that and such that and are associate by .

By now, we've reached a point that in the polynomial ring over a UFD , no matter the decomposition of content into irreducible elements or the decomposition of primitive part into irreducible primitive polynomials should be unique, and is indeed a UFD.

And from the conclusion we know polynomial rings like , , over UFD are all UFDs, although none of them is PID.

Conclusion

In this text, we've first given exact definition of unique factorization that a non-unit non-zero element is said to be uniquely factorized in the integral domain when any pair of its decomposition into irreducible elements have the same length while there's a way to match every pair of factor irreducible elements so that they're associate pairwisely. By digging into the essence of unique factorization, we've concluded that an integral domain that every non-unit non-zero can be uniquely factorized, which is known as the unique factorization domain (UFD), the factorization series of all non-unit non-zero element should have finite length, while all irreducible elements must be prime in the domain. These two criterion are sufficient and necessary for a integral domain to be UFD.

In order to fulfil the criterion, from the classical examples and we've learned before this text so far, we test the integral domain in which every ideal inside is principle, which is known as the principal ideal domain (PID). And in this domain, all principal ideals of irreducible elements are maximal and thus prime, while having an infinite ascending chain of principal ideal will surely violate the constraint that every ideal should be principal. So in this way, we've proved that all PIDs are UFDs.

On the other hand, we've also generalize the integral domain equipped with Euclidean algorithm and Bezout's identity into Euclidean domain. By putting constraint on the greatest common divisor every time we'll obtain by Euclidean algorithm, the Euclidean valuation, which is a way to "measure" the distance of a principal ideal (corresponding to the current greatest common divisor) to the whole ideal, we are able to trap the every ideal into a finite chain of principal ideal, and thus every ideal in Euclidean domain is principal, and it is indeed a UFD. Euclidean domain comes in handy for proving integral domains equipped with proper division and Euclidean algorithm to be UFDs, like the Gaussian integers and the Eisenstein integers .

Finally, we've inspected the polynomial rings over UFDs. In these rings, we are able to define similar concepts and conclusions to those in , like the primitive polynomial, Gauss's lemma and Eisenstein criterion. And most importantly, for any polynomial ring over a UFD , by utilizing the unique factorization property in and , fundamental properties of associate primitive polynomials in , and Gauss's lemma, we've successfully proved that is also a UFD although it's generally not a PID.

By now, we should be familiar with concepts and basic examples of unique factorization domains, which are foundations of the upcoming theories.

October 3, 2023