Suppose DL(P3(R),P2(R))D \in \mathcal L(\mathcal P_3(\mathbf R), \mathcal P_2(\mathbf R)) is the differentiation map defined by Dp=pDp = p'. Find a basis of P3(R)\mathcal P_3(\mathbf R) and a basis of P2(R)\mathbf P_2(\mathbf R) such that the matrix of DD with respect to these bases is

(100001000010).\begin{pmatrix} 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 \end{pmatrix}.

Consider the basis (x3,x2,x,1)(x^3, x^2, x, 1) of P3(R)\mathcal P_3(\mathbf R) and the basis (3x2,2x,1)(3x^2, 2x, 1) of P2(R)\mathcal P_2(\mathbf R). Clearly D(x3)=3x2D(x^3) = 3x^2 etc, meaning the matrix is as desired.


In more general terms, we picked (x3,x2,x,1)(x^3, x^2, x, 1) so that the nullspace was at the end, then picked (Dx3,Dx2,Dx)(Dx^3, Dx^2, Dx) as our output basis.

Matrix multiplication can be interpreted as writing our vectors in terms of a new basis, so picking a basis that makes that look like the identity by applying the transformation to each basis vector makes sense.

In the nice case where we're dealing with an invertible linear map TL(V,W)T \in \mathcal L(V,W) we can pick v1,,vnv_1,\dots,v_n a basis of VV then pick wj=Tvjw_j = Tv_j, the wjw_j's will be a basis for WW since TT is invertible meaning M(T)\mathcal M(T) with respect to these bases is the identity.