The Isomorphism L(V,W) ≅ Matrices

The isomorphism $L(V,W) \cong \mathbb{F}^{m\times n}$ is the bridge that makes numerical computation possible for abstract linear algebra problems: any software implementing operations on abstract linear maps between finite-dimensional spaces (e.g. differential operators restricted to polynomial spaces of bounded degree, or linear transformations of images and signals) works, under the hood, entirely with the matrix representative. In quantum mechanics, observables and finite-dimensional quantum states are manipulated as linear operators on Hilbert spaces, and this isomorphism is exactly what allows physicists to write down explicit matrices (Pauli matrices, Hamiltonians) for operators originally defined abstractly. In computer graphics, affine and linear transformations of 3D models are stored and composed as $4\times 4$ or $3\times 3$ matrices precisely because this theorem guarantees no information about the underlying linear map is lost in that representation. In statistics and machine learning, linear regression, PCA, and linear layers of neural networks are all linear maps between finite-dimensional spaces whose entire computational implementation is via the matrix $M(T)$; the dimension count $(\dim V)(\dim W)$ literally tells you the number of trainable parameters in a linear layer. In representation theory, this isomorphism (applied fiber-by-fiber, or representation-by-representation) underlies the identification of group representations with matrix groups, connecting abstract symmetry to concrete computable matrices, feeding directly into Matrix Multiplication as Composition once one studies how composed linear maps correspond to matrix products.

:::quiz[What breaks in the theorem if V is infinite-dimensional, as in the counterexample with $\mathbb{R}[x]$?]

• Nothing; the theorem still holds verbatim
• M(T) can no longer be represented by a finite matrix at all ✓
• L(V,W) becomes zero-dimensional
• W must also become infinite-dimensional

Ask the tutor: If I change the codomain basis $\gamma$ to $\gamma'$, exactly how does the new matrix $M'(T)$ relate to the old matrix $M(T)$?

Think of $L(V,W)$ as a warehouse of transformations — rules that eat a vector in $V$ and produce a vector in $W$ — while $\mathbb{F}^{m\times n}$ is a warehouse of grids of numbers. A priori these look like different kinds of objects entirely: one is defined by a universal property (linearity of a function), the other by finite data (an array). The bridge between them is a choice of bases, which lets us describe every vector in $V$ by $n$ coordinates and every vector in $W$ by $m$ coordinates. Once that translation dictionary is fixed, a linear map $T$ is completely pinned down by where it sends the $n$ basis vectors $v_1,\dots,v_n$ — and each image $Tv_j$ is completely pinned down by its $m$ coordinates relative to $\gamma$. So the "shape" of the information needed to specify $T$ is exactly an $m\times n$ grid: $m$ numbers for each of $n$ basis vectors. This is why the theorem is not a coincidence — it reflects the fact that both $L(V,W)$ and $\mathbb{F}^{m\times n}$ are, in an unavoidable sense, parametrized by the same $mn$ independent scalar choices. The picture below shows this translation for a concrete map $T: \mathbb{R}^3 \to \mathbb{R}^2$: each basis vector of the domain (3 of them) maps to a vector in the codomain, recorded as a column of 2 numbers, giving a $2\times 3$ grid of $6 = 2\cdot 3$ total scalars.

Example 1 (an explicit non-square, non-symmetric case). Let $V = \mathcal{P}_2(\mathbb{R})$ (polynomials of degree $\le 2$) with basis $\beta = (1, x, x^2)$, so $n = \dim V = 3$. Let $W = \mathbb{R}^2$ with the standard basis $\gamma = (e_1, e_2)$, so $m = \dim W = 2$. Consider the linear map

i.e., $T$ sends a polynomial to the pair (value at $1$, derivative at $0$). This is linear because evaluation and differentiation are each linear, and a pair of linear functionals assembles into a linear map into $\mathbb{R}^2$. Compute the images of the basis vectors:

(Check: for $p=1$, $p(1)=1, p'(0)=0$; for $p=x$, $p(1)=1, p'=1$ so $p'(0)=1$; for $p=x^2$, $p(1)=1$, $p'=2x$ so $p'(0)=0$.) Writing these as coordinate columns in $\gamma$ (which is just the standard basis, so coordinates equal the vectors themselves) gives the matrix

This is a genuine $mn = 2\cdot 3 = 6$-parameter object living inside the $6$-dimensional space $L(\mathcal{P}_2,\mathbb{R}^2) \cong \mathbb{R}^{2\times 3}$, exactly as the theorem predicts. Now let's verify additivity concretely. Let $S(p) = (p(0), p'(1))$, another linear map $\mathcal{P}_2 \to \mathbb{R}^2$. Compute:

(Check: $p=x^2 \Rightarrow p(0) = 0$, $p'=2x \Rightarrow p'(1) = 2$.) So

Now $(S+T)(p) = (p(1)+p(0),\ p'(0)+p'(1))$. Directly:

so

confirming linearity of $M$ (part 1 of the theorem) in this concrete instance.

Example 2 (surjectivity: reconstructing a map from an arbitrary matrix, asymmetric bases). Let $V = \mathbb{R}^2$ with the non-standard basis $\beta = (v_1,v_2) = \big((1,1),\,(1,-1)\big)$, and let $W = \mathbb{R}^3$ with the standard basis $\gamma = (e_1,e_2,e_3)$. Here $n=2$, $m=3$. Take the arbitrary matrix

Following the surjectivity proof, define $T$ by its action on $\beta$:

To describe $T$ on an arbitrary input $(a,b) \in \mathbb{R}^2$ (standard coordinates), we must first express $(a,b)$ in the basis $\beta$. Solve $(a,b) = c_1(1,1) + c_2(1,-1)$: adding and subtracting gives $c_1 = \tfrac{a+b}{2}$, $c_2 = \tfrac{a-b}{2}$. Then by linearity,

Expanding each coordinate:

(Check the middle coordinate: $-\tfrac{a+b}{2} + \tfrac{3a-3b}{2} = \tfrac{-a-b+3a-3b}{2} = \tfrac{2a-4b}{2} = a-2b$. ✓.) Let's verify $M(T) = A$ directly from this formula. Compute $Tv_1 = T(1,1)$: plug $a=1,b=1$: $(1+1,\ 1-2,\ \tfrac{5}{2}-\tfrac{5}{2}) = (2,-1,0)$. ✓ matches column 1 of $A$. Compute $Tv_2 = T(1,-1)$: $a=1,b=-1$: $(1-1,\ 1+2,\ \tfrac52+\tfrac52) = (0,3,5)$. ✓ matches column 2 of $A$. This confirms concretely that the linear map $T$ constructed from $A$ really does satisfy $M(T)=A$, illustrating surjectivity (part 3) with a genuinely non-standard domain basis.