Vector Spaces and Linear Transformations


We've actually informally introduced and discussed about vector spaces when we were proving the Wedderburn's little theorem, stating that finite division rings are vector spaces over their centers. We are using vector space as a tool for deriving numberical relationship between the elements in the division ring and and the elements in the center.

Vector spaces and linear transformations are concepts from undergraduate linear algebra, but in this text, we will extend them to ring theory discuss them in depth: first we will give vector spaces a formal and rigid definition and formalize them as algebraic objects; then we will discuss their homomorphisms, which are called the the linear transformations; finally we will reveal and prove that all linear transformations on finite dimensional vector spaces are representable by matrices, from a more abstract point of view and with more details.

Vector spaces

Vector spaces in ring theory is an extension of vector spaces of real numbers (like of -dimensional real vectors) to arbitrary field, that a vector space over field is an abelian group of vectors with vector addition and additive identity , equipped with scalar multiplication (where is called the scalar), so that 1:

  1. .
  2. .
  3. .
  4. .

By using somebasic trick in ring theory, it won't be hard to show that for any , we have , and .

Subspace and linear independence

A subspace is defined as a subgroup of closed under scalar multiplication, so that .

A way of constructing subspace of is to take some vectors out of , and make the set of linear combinations of them. Then there's by:

Such a subspace is called the spanning set of , or the subspace spanned by .

While there's no constraint on what vectors must be chosen to span the subspace , without losing generality, assume non-zero vector can be represented by other vectors through:

Clearly not all are zero, otherwise would be zero then. And For any vector with non-zero scalar , there's:

So is also a vector representable by other vectors. And for any vector that can be located by , obviously it can also be located by .

This is what will happen when there were "redundant" vectors chosen to span the subspace . And such phenomenon inspire us to define the linear (in)dependence for vectors, that among the vectors , if some vectors are representable by other vectors, then there must be:

And these vectors are said to be linearly dependent. Contrary to this, if all vectors are not representable by others remaining, in order for , the only possibility is . So we define these vectors to be linearly independent when there's:

So when linearly dependent vectors are chosen to span , we can safely remove some "redundant" vectors from them without affecting the remaining vectors' capability to span . And each time when we are able to remove "redundant" vector, by the rationale above there're always multiple choices (two or more) to remove, so there're many possibilities when we reach the point that remaining vectors are linearly independent and removing any more of them will fail to span . But for any two possibilities among them that and , which are both linearly independent and will span , there must be .

To prove, first let's show that if is spanned by linearly independent vectors , then vectors must be linearly dependent if .

Compatible with matrix multiplication there's:

Where .

Compatible with Gauss's elimination on the first -rows, where must be invertible. If after the elimination, is not upper-triangular. then the last row of must be all zeroes, and the last row of must also be . Let the last row of be , which must not be all-zeros, then there's:

So the first vectors of are linearly dependent, and we are done.

Otherwise If is upper-triangular, we can furtherly eliminate the last rows with Gauss's elimination , so that will be cleared to zeros after elimination, alongside with , that:

Noted that there's , and since there's , we have . So among the vectors , the last vectors are representable by first vectors, and they are thus linear dependent.

Then, if subspace is spanned by both and , where they are both linear indepdent but , without losing generality, let's assume it's the case that , then must be linear dependent, which is a contradiction.

Finite dimensional vector spaces and its subspaces

The set of linearly independent vectors that spans subspace is called the basis of . Obviously there're always multiple choices of : given a choice of basis , linearly combine them so that they are now linearly dependent, and this time keep the combined vector and choose another vector in for removal. But for any two choices of basis, the number of vectors inside must always be the same, and is truly an invariance of subspace spanned by basis. We call it the dimension of , and denoted it as .

Conversely, given a basis , we can use it to span a subspace, and such a subspace is denoted as . For convenience, if the basis is finite, then is said to be a finite dimensional subspace, or subspace of dimension . And we can treat the vector space itself as a trivial subspace, calling it finite dimensional vector space or vector space of dimension correspondingly. Specially, there's .

Any subspace of finite dimensional vector space is also finite dimensional. This can be shown by performing the algorithm we've described here, by starting at , iteratively selecting random vector , and adding it to . The randomly selected vector is guaranteed to be linearly independent with vectors in already selected basis , otherwise there's , and , which is a contradiction. The algorithm guarantees that , and will select no more than vectors, otherwise they will be linearly dependent, so algorithm is guaranteed to halt. However the basis of are not always the basis of : consider the -dimensional Cartesian coordinate system with basis corresponding to -axis, the vectors on the line is a subspace of spanned by , but not possibly by any one from .

Linear transformations

Just like other algebraic objects, we would like to define their "homomorphisms". And for vector spaces, their homomorphisms are conventionally called the linear transformations. That is, for two vector spaces over , a linear transformation is a map fulfilling:

So is a group homomorphism between abelian groups in the first place, so its kernel is closed under addition. And in order for to be the kernel of a linear transformation, it must fulfil , which means it must also be closed under scalar multiplication. Given that is closed under addition and scalar multiplication, it must be subspace of .

Conversely, Consider the canonical homomorphism of any subspace . When viewing as quotient group, there's ; while it's clear that , so the quotient group is compatible with homomorphic scalar multiplication. By now, we know the canonical homomorphism of any subspace is a linear transformation.

Rank-nullity theorem

The kernel is conventionally called the null space of , while the image is called the range space of . For finite dimensional vector space , there's rank-nullity thereom stating that , where the dimension of null space is called the nullity and the dimension of the range space is called the rank.

To prove, since is a subspace of , it is also finite dimensional, let one of its basis be , since vectors in are already linearly independent and is subspace of , we can continue on selecting more vectors from so that is one of the basis of , and maps all vectors in by .

To show the set that set of vectors are linearly independent, assume they are not linearly independent instead, so we have , for some s that are not all zeroes, which in turns implies for some s that are not all zeroes there're and . In this way, the vectors in are not linearly independent, which is a contradiction. So we have , and by putting all information above together, we have , and we are done.

For linear transformation between finite dimensional vector spaces that , we would like to show that is injective iff it's surjective: when is surjective, and , so it's injective; and when is injective, let the the basis of be , for any vector , it's linearly independent with vectors in , and now we have linearly independent vectors in , which is a contradiction. So must be surjective as long as it's injective.

Quotient space and the first isomorphism theorem

On the other hand, consider the vectors , which are extended from the basis of to span the whole , after being attached to as cosets , can span all cosets in the quotient space : for every vector , by performing the canonical homomorphism of , we will see the portion of absorbed into , leaving identifying the coset in . Therefore, by using as basis for , we can see they span all cosets in . Each base vector for corresponds to the base vector for , fulfilling the first isomorphism theorem unexceptionally.

Isomorphisms of vector spaces of the same dimension

Let be set of the column vector of elements from in the form of . Compatible with vector addition and scalar multiplication, is indeed a vector space of dimension , with vectors in basis . Any vector space of dimension is isomorphic to , since given its basis , consider the linear transformation defined by , it's obviously injective otherwise vectors in will be linearly dependent, and is thus surjective. So linear transformation is an isomorphism between and of dimension .

Sum and intersection of subspaces

For subspaces of vector space , when viewing as abelian groups, it's obvious that and are both normal subgroups of . And by , , both and are subspaces of . When is trivial, is inner direct product of and and is called the direct sum, denoted as conventionally. When viewing as normal subgroups, they obeys the second isomorphism theorem that . Let be finite dimensional vector space, consider the canonical homomorphism of and , we have , and thus .

Matrices of linear transformations

We are almost familiar with linear algebra, and must have heard (I think those who are not major in mathematics are generally on the level of "heard") the statement that all linear transformations of finite dimensional vector space are matrices, but we may not know how they are actually associated and proved they are truly the same. And let's focus on the connection between matrices and linear transformations now.

All linear transformations on finite vector spaces are representable by matrices

Let and be vector spaces of dimension and and with basis and . Consider a linear transformationation , any vector will be transformed into . For every vectors , if we know can be represented by vectors in that , then there's . And compatible with matrix multiplication there's:

By now, we can clearly see the role matrices are playing and how they are associated with linear transformations: for a linear transformation between finite dimensional vector spaces, once we are able to determine basis of and of , we will know how to transform their scalars or coordinate, which can be represented as matrix multiplication. And the matrix for transforming scalars or coordinate is called the matrix of relative to basis of and basis of .

Let there be two linear transformation and with basis for , basis for and basis for respectively, matrix be matrix of relative to and , and matrix be matrix of relative to and . Given that:

The matrix of is multiplication of 's own matrices when fixing the source, intermediate and target vector spaces' basis.

This leads us to that all matrix of bijective linear transformations are invertible: for a matrix of a bijective linear transformation relative to basis and , consider the matrix of relative to basis and . Given that the composition is the identity linear transformation on , whose matrix relative to must only be the identity matrix , there must only be , so must be invertible with inverse , which can be evaluated by inspecting how each vector in can be represented as linear combination of vectors in .

Vector spaces defined by matrix and their relationship with linear transformation

Similarily we define the null space of matrix as . where is matrix of relative to basis of and basis of . The linear transformation defined by is an isomorphism, and we will show :

  1. , , .
  2. .
  3. .

To soothe our intuition, we could define some kind of "inner product" between vectors as , and two vectors are said to be orthogonal or perpendicular when there's , then clearly are all vectors orthogonal to every row vectors of .

Define the row space of matrix as the subspace of spanned by row vectors , then by , any vector from the row space of is orthogonal to any vector from the null space of .

On the other hand, define the column space of matrix as the subspace of spanned by its column vectors , and the column rank of matrix as the dimension of the column space. Naturally the map defined by is a linear transformation with its image being . If is the matrix of linear transformation relative to basis of and basis of , we've prove that , and from this relationship we can decompose into three linear transformations: first we extract the coordinate with respect to basis with linear transformation defined by , it's clear that is bijective; then we apply the linear transformation naturally defined by ; finally we reassemble the coordinate with respect to basis with linear transformation defined by , it's also bijective for domain and range . So there's , while are both bijective, it's only possible that . In this way we build up the relationship between the column rank of and the rank of linear transform .

Row rank equals column rank in a matrix

Given that there's column rank, there's also row rank that is defined to be the dimension of the row space. And for matrix , there's always :

  1. First let's do the Gauss's elimination on the rows of , so that is in upper-triangle or staircase shape. When viewing form the row side, the row vectors are transformed into where . And for each non-zero row vector whose -th column is not zero (and there's by it's in upper-triangle or staircase case), it is not representable by since they have non-zero component on the first columns; it is also not representable by since their -th column is zero. So the number of non-zero row vectors in is directly . Given that , defines a bijective linear transformation between and , and we have .
  2. When viewing from the column side, the column space of are spanned by vectors . Given that and , is bijectively transformed from by and we have .
  3. Let's furtherly perform the Gauss's elimination on the columns of , so that it becomes . Given that , defines a bijective linear transformation between and . By joining all these above together, we have , and we are done.

And given that there's for any matrix , we define the rank of matrix to be the column rank or the row rank of , denoting it as . In this way, the rank of the matrix is in accordance with the rank of the linear transformation it represents.

Conclusion

In this text, we introduce the concept of vector spaces and linear transformation from a more abstract point of view, by studying them as algebraic objects and their homomorphisms between the objects.

For vector spaces, the most special kind of them is the finite dimensional vector space, in which we can take finite linear independent vectors to span the whole space. There're always multiple choices of basis but the dimension of the space is an invariance, constraining the exact number of linear independent vectors required to span the space. Every subspace of a finite dimensional vector space can also be shown to be finite dimensional.

For linear transformations, which is the homomorphism between vector spaces, it can be shown any subspace of the domain vector space can be used as kernel as a linear transformation. For finite dimensional vector spaces, the rank-nullity theorem must hold, and all -dimensional vector spaces are isomorphic to where is the field of scalars.

Finally, we've revealed the interconnection between matrices and linear transformations, that when fixing sets of basis of the domain and range vector spaces, linear transformations can always be represented by matrices. We've also inspected the crucial vector spaces defined by the matrix and related to the linear transformations: the row space of the matrix is the vector space spanned by row vectors of the matrix; the null space of the matrix is formed by the vectors perpendicular to the row space, it has the same dimension as the null space of the linear transformation; the column space of the matrix is the vector space spanned by column vectors of the matrix, it has the same dimension as the rank of the linear transformation. The dimension of the row space, which is called the row rank, can be shown to be equal to the dimension of the column space, which is called the column rank. Given that there's such accordance, we call the row rank and the column rank of the matrix the rank of the matrix uniformly, and it equals to the rank of the linear transformation.

Although what's addressed in this text is far from what one should know about linear algebra, but I believe it has introduced some fundamental concepts and can make one's mind clear about linear algebra: it's a branch of mathematics studying vector spaces and linear transformations, rather than merely telling undergraduates how to manipulate matrices, evaluating determinants, evaluating eigenvalues and eigen vectors, and so on. It's also enough for us to "toolizing" vector spaces and linear transformations in ring theory for now.


  1. It may look like some kind of 's "acting" on group , although there's nothing called "ring action", but instead something called "-module" defined for rings in commutative algebra, and actually all -modules for field are vector spaces. We will not discuss it in depth in the current text. [return]
September 7, 2023