In this lecture prof. Kutz gives the Dynamic Mode Decomposition modes:

$\Phi = X' V_r \Sigma_r^{-1} W $

which are the eigenvectors of the linear propagator matrix. This results from splitting the full data matrix $X$ into $X' = X(:,2:\text{end})$ and $X_1 = X(:,1:\text{end}-1)$. $U_r$, $\Sigma_r$ and $V_r$ are matrices resulting from the SVD of the data matrix $X_1$ and rank-$r$ truncation afterwards. $W$ is the approximation to the eigenvectors of the linear propagator matrix.

However, in the textbook by prof. Kutz Data-driven modeling and scientific computation, this formula is given for the DMD modes:

$\Phi = U_r W$

and this is also the formulation that was originally given by Peter Schmid.

I don't see them being equivalent equations, but perhaps I overlook something. To my understanding, they would be equivalent if in the first equation we used the rank-$r$ approximation matrix $X_{1 r} \approx X_1$ instead of $X'$:

$X_{1 r} V_r \Sigma_r^{-1} W = U_r \Sigma_r V_r^T V_r \Sigma_r^{-1} W = U_r W$

since $V_r$ is both orthogonal and orthonormal, and $\Sigma_r$ is a diagonal matrix.

Could you please explain the logic/intuition behind both approaches and if and how these two approaches differ from each other?


There's two issues here. First, are we using X(:,2:end) or X(:,1:end−1)? I have no idea.

Second, are we using the rank r approximation or the full matrix? Turns out it doesn't matter. The missing bit is actually 0. Let $USV^T$ be the full SVD of $X_1$. Let $S_r$, $U_r$, and $V_r$ be identical to $S$, $U$, and $V$ except that columns $r+1$ and onwards are set to 0. So, $U_rS_rV_r^T$ is a truncated SVD, but the individual matrices all stay at their original (non-truncated) dimensions. This allows me to show how they interact with one another. Let $S_{r+1}$, $U_{r+1}$ and $V_{r+1}$ be the remainders: they equal $S$, $U$ and $V$ respectively but with the first $r$ columns set to 0. By looking back at the outer product form of the SVD:

$$\sum_r u_r s_r v_r^T$$

... you'll hopefully see that

$$X_1 = U_{r} S_{r}V_{r}^T + U_{r+1} S_{r+1}V_{r+1}^T= X_{1r} + U_{r+1} S_{r+1}V_{r+1}^T$$.

The key to seeing why

$$X_{1}V_{r} = X_{1r}V_{r}$$

is that the remainder term, ending in $V_{r+1}^T$, will completely cancel with $V_r$.


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.