Given an irreducible non-symmetric, non-negative real matrix $A$ with eigenvalues $\lambda_1 \geq \lambda_2 \geq \lambda_3 \geq \cdots$ and $k$ of its largest right eigenvectors:

\begin{eqnarray} % A v_1 &= \lambda_1 v_1 \\ A v_2 &= \lambda_2 v_2 \\ \vdots & \vdots \\ A v_k &= \lambda_k v_k \\ % \end{eqnarray}

What (if anything) can we say about it's $k$ largest left eigenvectors? \begin{eqnarray} % w_1 A &= \lambda_1 w_1 \\ w_2 A &= \lambda_2 w_2 \\ \vdots & \vdots \\ w_k A &= \lambda_k w_k \\ % \end{eqnarray}


If $wA = \lambda w$, we have, using transposition, $A^T w^T = \lambda w^T$.

Assuming $A$ is diagonalizable, we have $A = T^{-1} D T$ for an invertible matrix $T$ and a diagonal matrix $D$, so $A^T = T^T D^T (T^{-1})^T = T^T D (T^T)^{-1}$.

Therefore, $\lambda w^T = A^T w^T = T^T D (T^T)^{-1} w^T$, or equivalently, $\lambda (T^T)^{-1} w^T = (T^T)^{-1}A^T w^T = D (T^T)^{-1} w^T$. But $\lambda v = D v$ iff $T^{-1} v$ is an eigenvector of $A$ (simple calculation), so $T^{-1} (T^T)^{-1} w^T = (T^T T)^{-1} w^T$ is an eigenvector of $A$.

Thus, $(T^T T)^{-1} w^T = v$ for some right eigenvector $v$, or equivalently, $w = ((T^T T) v)^T = v^T T^T T$. This should generalize to the non-diagonalizable case using arguments of the Jordan form.

| cite | improve this answer | |
  • $\begingroup$ Thanks for the answer Johannes. As I see it, doesn't the term $T^T T$ require to know all of the right eigenvectors $v$? Can you calculate a subset of the $w$'s from a subset set of the $v$'s? In the question as stated, I only know a small subset of the $v$'s ($k$ largest). $\endgroup$ – Hooked Apr 19 '12 at 16:22
  • $\begingroup$ It can probably be done by doing something smart with block matrices; just play around a bit. $\endgroup$ – Johannes Kloos Apr 19 '12 at 16:37
  • $\begingroup$ Actually, scratch that. It could well be that there is no good solution if you have too little information about the eigenvectors; maybe there is some numeric trickery that helps approximate $T^T T$? $\endgroup$ – Johannes Kloos Apr 19 '12 at 16:39
  • $\begingroup$ If you come across it, I'd be interested. Either way, your answer, and more importantly, your quick proof was very useful. $\endgroup$ – Hooked Apr 19 '12 at 18:08

Your left eigenvectors are characterized by the null space of $V$, as explaned in the following.

The matrix form of your known equations is $$AV = VD$$ where $D = diag(\lambda_i)$ for $1 \le i \le k$ and the columns of $V$ are the known eigenvectors.

Writing in terms also of the column matrix of unknown vectors, call it $W$, we have $$A\pmatrix{V & W} = \pmatrix{V & W}\pmatrix{D & \mathbf{0} \\ \mathbf{0} & D_w}$$

Since $\pmatrix{V & W}$ is invertible (it is a basis) it diagonalizes A, and we may write $$\pmatrix{V & W}^{-1}A\pmatrix{V & W} = \pmatrix{D & \mathbf{0} \\ \mathbf{0} & D_w}$$

Denote $\pmatrix{V & W}^{-1}=\pmatrix{\tilde{V}^T \\ \tilde{W}^T}$ to obtain (using $n$ as the dimension of $A$) $$\pmatrix{\tilde{V}^T \\ \tilde{W}^T}\pmatrix{V & W} = \pmatrix{\mathbf{I}_k & \mathbf{0} \\ \mathbf{0} & \mathbf{I}_{n-k}} =\pmatrix{\tilde{V}^T V & \tilde{V}^TW \\ \tilde{W}^TV & \tilde{W}^T W}\tag{1}$$ In terms of the $k$ known right eigenvectors $V$: $$\pmatrix{\tilde{V}^T \\ \tilde{W}^T} V = \pmatrix{\mathbf{I}_k \\ \mathbf{0}}$$ From this we see that $n-k$ of the left eigenvectors $\tilde{W}^T$ are contained in the perpendicular space to $V$ since $\tilde{W}^TV = \mathbf{0}$.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.