PropositionSplitting by eigenspaces

Suppose \(\EigenVal{\lambda}_1, \dots ,\EigenVal{\lambda}_k\) are distinct eigenvalues of an \((n,n)\)-matrix \(\Mtrx{A}\). Let \(V\) be the subspace of \(\RNrSpc{n}\) spanned by the associated eigenspaces \(\EigenSpc{\EigenVal{\lambda}_1}, \dots ,\EigenSpc{\EigenVal{\lambda}_k}\). Then

\[V = \EigenSpc{\EigenVal{\lambda}_1} \dotplus \cdots \dotplus \EigenSpc{\EigenVal{\lambda}_k}\]

is the internal sum of these eigenspaces. In particular,

\[\DimOf{V} = \GmtrcMltplcty{\EigenVal{\lambda}_1} + \cdots + \GmtrcMltplcty{\EigenVal{\lambda}_k}\]

Proof

By construction, \(V\) is the sum of its the given eigenspaces. It, therefore, only remains to show that these eigenspaces only have the 0-vector in common. In other words, if \(\Vect{a}_i\) is in \(\EigenSpc{\EigenVal{\lambda}_i}\), we need to show that

\[\Vect{a}_i + \cdots +\Vect{a}_k = \Vect{0} \quad\text{if and only if} \quad \Vect{a}_1=\cdots = \Vect{a}_k=\Vect{0}\]

Only the direction from left to right presents a challenge, and here we argue by induction on \(k\). For \(k=1\), the claim is a tautology. Now suppose the claim is true for \(1\leq k-1\). We need to infer its truth for \(k\). So suppose

\[\Vect{a}_1 + \cdots +\Vect{a}_k = \Vect{0}\]

Then

\(A(\Vect{a}_1 + \cdots +\Vect{a}_k)\)\(=\)\(\Vect{0}\)
\(\EigenVal{\lambda}_1\cdot \Vect{a}_1 + \cdots + \EigenVal{\lambda}_k\cdot \Vect{a}_k\)\(= \)\(\Vect{0}\)

Now we need the assumption that the eigenvalues \(\EigenVal{\lambda}_1, \dots , \EigenVal{\lambda}_k\) are distinct. Therefore, at least one of them is not 0. By rearranging indices, if necessary, we may achieve \(\EigenVal{\lambda}_k\neq 0\). Dividing the last equation by \(\EigenVal{\lambda}_k\), and recalling the original equation yields

\(\Vect{a}_1 + \cdots + \Vect{a}_{k-1} + \Vect{a}_k\)\(=\)\(\Vect{0}\)
\(\tfrac{\EigenVal{\lambda}_1}{\EigenVal{\lambda}_k}\cdot \Vect{a}_1 + \cdots + \tfrac{\EigenVal{\lambda}_{k-1}}{\EigenVal{\lambda}_k}\cdot \Vect{a}_{k-1} + \Vect{a}_k\)\(=\)\(\Vect{0}\)

Subtracting the first equation from the second yields

\((\tfrac{\EigenVal{\lambda}_1}{\EigenVal{\lambda}_k}-1)\cdot \Vect{a}_1 + \cdots + (\tfrac{\EigenVal{\lambda}_{k-1}}{\EigenVal{\lambda}_k}-1)\cdot \Vect{a}_{k-1}\)\(=\)\(\Vect{0}\)

In this last equation, each vector \((\tfrac{\EigenVal{\lambda}_i}{\EigenVal{\lambda}_k}-1)\cdot \Vect{a}_i\) belongs to \(\EigenSpc{\EigenVal{\lambda}_i}\). By induction hypothesis,

\((\tfrac{\EigenVal{\lambda}_i}{\EigenVal{\lambda}_k}-1)\cdot \Vect{a}_i\)\(=\)\(\Vect{0}\)

We also know that \(\EigenVal{\lambda}_i\neq \EigenVal{\lambda}_k\), for \(1\leq i\leq k-1\). Therefore

\(\tfrac{\EigenVal{\lambda}_i}{\EigenVal{\lambda}_k} -1\)\(\neq\)\(0\)

for \(1\leq i\leq k-1\). But then \(\Vect{a}_i=\Vect{0}\) for \(1\leq i\leq k-1\). The original equation now yields \(\Vect{a}_k=\Vect{0}\), and this implies the claim.