Operators on Inner Product Space¶
Orthogonal Complements¶
Theorem for null space and range of \(T^*\)
Suppose \(T\in \mathcal{L}(V, W)\), then
(i) \(N T^*=(R T)^\perp\).
(ii) \(R T^*=(NT)^\perp\).
\(\square\)
Singular value decomposition¶
For a linear map \(T \in \mathcal{L}(V,W)\), we could decompose it as we have for self-adjoint operator or normal operator.
Recall the important Riesz representation theorem in inner product space.
Riesz representation theorem
Assume \(V\) is finite-dimensional and \(\phi\) is a linear functional on \(V\), then there exists a unique vector \(v\in V\) such that
\(\square\)
In functional analysis, we have actually a similar result for infinite-dimensional spaces.
The following lemma of \(T^*T\) is necessary.
Lemma
Properties of \(T^*T\). Suppose \(T \in \mathcal{L}(V,W)\)
(i) \(T^*T\) is a self-adjoint operator on \(V\). We could also check \(TT^*\) is a self-adjoint operator on \(W\).
(ii) \(N T^*T=N T\).
(iii) \(R T^*T = R T^*\).
(iv) dimension. \(\text{Dim} R T=\text{Dim} R T^*=\text{Dim}R T^*\).
(i) by definition.
(ii) \(N T\subset N T^*T\) is apparent. Assume \(v\in N T^*T\), \(T^*T v=0\), so \(\langle v, T^*Tv\rangle=0\), so \(\langle Tv, Tv\rangle=\|Tv\|=0\), which means \(Tv=0\).
(iii) \(R T^*T \subset R T^*T\) is apparent. For another direction, we use (ii) \(R T^*T=(N T^*T)^\perp=(N T)^\perp=R T^*\).
(iv) Use fundamental theorem of linear maps.
Definition of singular value
Assume a linear operator \(T\in \mathcal{L}(V, W)\), the singular values of \(T\) are defined as the nonnegative square roots of the eigenvalues of \(T^*T\), listed in decreasing order.
(SVD) Singular value decomposition
Assume a linear operator \(T\in \mathcal{L}(V, W)\), with its positive singular values \(s_1,\cdots, s_r\). Then there exists orthonormal lists \(e_1,\cdots, e_r\subset v\), \(f_1,\cdots, f_r\subset W\), such that
Here we denote that \(V\) and \(W\) is finite-dimensional. And the proof is constructive. This method also gives info about the eigenvectors construction.
Let \(s_1,\cdots,s_n\) to be the singular value of \(T(\text{Dim}V=n)\), where \(s_{r+1},\cdots, s_n\) are zero singular values.
- Apply spectral thoerem to \(T^*T\) and there exists orthonormal basis \(e_1,\cdots, e_n\subset V\), such that
- Define \(f_k=\frac{Te_k}{s_k}\) for \(k=1,\cdots, r\).
this is actually orthonormal basis in \(W\). This is also inspired by \(T=W\Sigma V^T\), so \(TV=W\Sigma\), so \(TV\Sigma^{-1}=W\), which shows a relationship of basis from \(V\) to \(W\) space.
- Prove the proposition by expressing \(v\) in the constructed orthonormal basis
for \(k\geq r\), \(Te_k=0\) because \(T^*T e_k =0\cdot e_k\) and Property of self-adjoint \(T^*T\) (ii).
We could also check that the matrix with respect to basis \(\{e_k\}_{1\leq k\leq r}\) and \(\{f_k\}_{1\leq k\leq r}\) which should be extended.
Note we have \(\{e_k\}_{1\leq k\leq n}\), and from the above proof we have \(Te_k=s_k f_k\) for \(k\leq r\) and \(0\) for \(k >r\). We shall extend \(\{f_k\}_{1\leq k\leq r}\) to \(\{f_k\}_{1\leq k\leq m} (\text{Dim} W=m)\) by utilizing \(N T^*\). This is because we want to solve \(R(T)^\perp\), which equals \(NT^*\) by Theorem for null space and range of \(T^*\). (Readers should double check the dimension of \(N T^*\), which is \(m-r\), for \(\text{Dim} RT=r\).)
\(\square\)
Matrix version of SVD, a compact SVD form
Assume \(A\) is an \(m\)-by-\(n\) matrix of rank \(r\geq 1\). Then there exists an \(m\)-by-\(r\) matrix \(W\) with orthogonal columns, an \(r\)-by-\(r\) diagonal matrix \(\Sigma\) with positive numbers on the diagonal, and an \(n\)-by-\(r\) matrix \(V\) with orthonormal columns such that
Let \(T: \mathbb{F}^n\rightarrow \mathbb{F}^m\) whose matrix with respect to the standard basis equals \(A\). From the above proof of the SVD theorem, we have \(\text{Dim}RT=r\) and
we make use of the above structure. Let
\(W\) to be the \(m\)-by-\(r\) matrix whose columns are \(f_1,\cdots,f_r\),
\(\Sigma\) to be the \(r\)-by-\(r\) diagonal matrix \(\Sigma\) with entries \(s_1,\cdots,s_r\),
\(V\) to be the \(n\)-by-\(r\) matrix whose columns are \(e_1,\cdots,e_r\).
Choose \(u_k\), a standard base of \(\mathbb{F}^m\), then apply this matrix
so \(AV=W\Sigma\), multiply both sides by \(V^*\) and we have \(A=W\Sigma V^*\). But we have to be careful.
Here actually \(VV^*= I\) does not hold absolutely. We have to argue as follows. If \(k\leq r\), \(V^*e_k=u_k\), so \(VV^*e_k=e_k\). Thus \(AVV^*v=Av\) for all \(v\in \text{span}(e_1,\cdots,e_m)\). For \(v\in \text{span}(e_1,\cdots,e_m)^\perp\), we have \(Av=0\) and \(V^*v=0\), so we also have \(AVV^*v=Av=0\).
\(\square\)