7. Singular Value Decomposition (SVD)¶

In addition to regular packages contained in Anaconda by default, this lecture also requires:

!pip install quandl

Collecting quandl
  Downloading Quandl-3.7.0-py2.py3-none-any.whl (26 kB)

Collecting more-itertools

  Downloading more_itertools-8.14.0-py3-none-any.whl (52 kB)
?25l
     |██████▎                         | 10 kB 34.6 MB/s eta 0:00:01
     |████████████▋                   | 20 kB 41.8 MB/s eta 0:00:01
     |██████████████████▉             | 30 kB 50.0 MB/s eta 0:00:01
     |█████████████████████████▏      | 40 kB 48.1 MB/s eta 0:00:01
     |███████████████████████████████▍| 51 kB 51.4 MB/s eta 0:00:01
     |████████████████████████████████| 52 kB 2.6 MB/s 
?25hRequirement already satisfied: six in /usr/share/miniconda3/envs/quantecon/lib/python3.9/site-packages (from quandl) (1.16.0)
Requirement already satisfied: python-dateutil in /usr/share/miniconda3/envs/quantecon/lib/python3.9/site-packages (from quandl) (2.8.2)
Requirement already satisfied: pandas>=0.14 in /usr/share/miniconda3/envs/quantecon/lib/python3.9/site-packages (from quandl) (1.4.2)
Requirement already satisfied: numpy>=1.8 in /usr/share/miniconda3/envs/quantecon/lib/python3.9/site-packages (from quandl) (1.22.4)
Requirement already satisfied: inflection>=0.3.1 in /usr/share/miniconda3/envs/quantecon/lib/python3.9/site-packages (from quandl) (0.5.1)
Requirement already satisfied: requests>=2.7.0 in /usr/share/miniconda3/envs/quantecon/lib/python3.9/site-packages (from quandl) (2.27.1)
Requirement already satisfied: pytz>=2020.1 in /usr/share/miniconda3/envs/quantecon/lib/python3.9/site-packages (from pandas>=0.14->quandl) (2021.3)
Requirement already satisfied: certifi>=2017.4.17 in /usr/share/miniconda3/envs/quantecon/lib/python3.9/site-packages (from requests>=2.7.0->quandl) (2021.10.8)
Requirement already satisfied: charset-normalizer~=2.0.0 in /usr/share/miniconda3/envs/quantecon/lib/python3.9/site-packages (from requests>=2.7.0->quandl) (2.0.4)

Requirement already satisfied: urllib3<1.27,>=1.21.1 in /usr/share/miniconda3/envs/quantecon/lib/python3.9/site-packages (from requests>=2.7.0->quandl) (1.26.9)
Requirement already satisfied: idna<4,>=2.5 in /usr/share/miniconda3/envs/quantecon/lib/python3.9/site-packages (from requests>=2.7.0->quandl) (3.3)

Installing collected packages: more-itertools, quandl

Successfully installed more-itertools-8.14.0 quandl-3.7.0

import numpy as np
import numpy.linalg as LA
import matplotlib.pyplot as plt
%matplotlib inline
import quandl as ql
import pandas as pd

7.1. Overview¶

The singular value decomposition is a work-horse in applications of least squares projection that form foundations for important machine learning methods.

This lecture describes the singular value decomposition and two of its uses:

principal components analysis (PCA)
dynamic mode decomposition (DMD)

Each of these can be thought of as a data-reduction procedure designed to capture salient patterns by projecting data onto a limited set of factors.

7.2. The Setup¶

Let $X$ be an $m \times n$ matrix of rank $p$.

Necessarily, $p \leq \min(m,n)$.

In this lecture, we’ll think of $X$ as a matrix of data.

each column is an individual – a time period or person, depending on the application
each row is a random variable describing an attribute of a time period or a person, depending on the application

We’ll be interested in two cases

A short and fat case in which $m << n$, so that there are many more columns (individuals) than rows (attributes).
A tall and skinny case in which $m >> n$, so that there are many more rows (attributes) than columns (individuals).

We’ll apply a singular value decomposition of $X$ in both situations.

In the first case in which there are many more individuals $n$ than attributes $m$, we learn sample moments of a joint distribution by taking averages across observations of functions of the observations.

In this $ m < < n$ case, we’ll look for patterns by using a singular value decomposition to do a principal components analysis (PCA).

In the $m > > n$ case in which there are many more attributes $m$ than individuals $n$, we’ll proceed in a different way.

We’ll again use a singular value decomposition, but now to construct a dynamic mode decomposition (DMD)

7.3. Singular Value Decomposition¶

A singular value decomposition of an $m \times n$ matrix $X$ of rank $p \leq \min(m,n)$ is

\[ X = U \Sigma V^T \]

where

\[ \begin{aligned} UU^T & = I & \quad U^T U = I \cr VV^T & = I & \quad V^T V = I \end{aligned} \]

where

$U$ is an $m \times m$ matrix whose columns are eigenvectors of $X^T X$
$V$ is an $n \times n$ matrix whose columns are eigenvectors of $X X^T$
$\Sigma$ is an $m \times n$ matrix in which the first $p$ places on its main diagonal are positive numbers $\sigma_1, \sigma_2, \ldots, \sigma_p$ called singular values; remaining entries of $\Sigma$ are all zero
The $p$ singular values are square roots of the eigenvalues of the $m \times m$ matrix $X X^T$ and the $n \times n$ matrix $X^T X$
When $U$ is a complex valued matrix, $U^T$ denotes the conjugate-transpose or Hermitian-transpose of $U$, meaning that $U_{ij}^T$ is the complex conjugate of $U_{ji}$.
Similarly, when $V$ is a complex valued matrix, $V^T$ denotes the conjugate-transpose or Hermitian-transpose of $V$

In what is called a full SVD, the shapes of $U$, $\Sigma$, and $V$ are $\left(m, m\right)$, $\left(m, n\right)$, $\left(n, n\right)$, respectively.

There is also an alternative shape convention called an economy or reduced SVD .

Thus, note that because we assume that $X$ has rank $p$, there are only $p$ nonzero singular values, where $p=\textrm{rank}(X)\leq\min\left(m, n\right)$.

A reduced SVD uses this fact to express $U$, $\Sigma$, and $V$ as matrices with shapes $\left(m, p\right)$, $\left(p, p\right)$, $\left( n, p\right)$.

7.4. Properties of Full and Reduced SVD’s¶

You can read about reduced and full SVD here https://numpy.org/doc/stable/reference/generated/numpy.linalg.svd.html

For a full SVD,

\[ \begin{aligned} UU^T & = I & \quad U^T U = I \cr VV^T & = I & \quad V^T V = I \end{aligned} \]

But these properties don’t hold for a reduced SVD.

Which properties hold depend on whether we are in a tall-skinny case or a short-fat case.

In a tall-skinny case in which $m > > n$, for a reduced SVD

\[ \begin{aligned} UU^T & \neq I & \quad U^T U = I \cr VV^T & = I & \quad V^T V = I \end{aligned} \]

In a short-fat case in which $m < < n$, for a reduced SVD

\[ \begin{aligned} UU^T & = I & \quad U^T U = I \cr VV^T & = I & \quad V^T V \neq I \end{aligned} \]

When we study Dynamic Mode Decomposition below, we shall want to remember this caveat because sometimes we’ll be using reduced SVD’s to compute key objects.

Let’s do an exercise to compare full and reduced SVD’s.

To review,

in a full SVD
- $U$ is $m \times m$
- $\Sigma$ is $m \times n$
- $V$ is $n \times n$
in a reduced SVD
- $U$ is $m \times p$
- $\Sigma$ is $p\times p$
- $V$ is $n \times p$

First, let’s study a case in which $m = 5 > n = 2$.

(This is a small example of the tall-skinny case that will concern us when we study Dynamic Mode Decompositions below.)

import numpy as np
X = np.random.rand(5,2)
U, S, V = np.linalg.svd(X,full_matrices=True)  # full SVD
Uhat, Shat, Vhat = np.linalg.svd(X,full_matrices=False) # economy SVD
print('U, S, V ='), U, S, V

U, S, V =

(None,
 array([[-0.55422319,  0.45878562, -0.63290178, -0.08137861, -0.27416284],
        [-0.35010242, -0.85724543, -0.16430617, -0.15976461, -0.30006221],
        [-0.64537478,  0.16626522,  0.73688466, -0.03085836, -0.1090689 ],
        [-0.13633697, -0.10083474, -0.06175662,  0.98267275, -0.04224893],
        [-0.3676528 , -0.12974691, -0.16008102, -0.03542345,  0.90615474]]),
 array([1.73650684, 0.31403256]),
 array([[-0.74885926, -0.66272906],
        [-0.66272906,  0.74885926]]))

print('Uhat, Shat, Vhat = '), Uhat, Shat, Vhat

Uhat, Shat, Vhat =

(None,
 array([[-0.55422319,  0.45878562],
        [-0.35010242, -0.85724543],
        [-0.64537478,  0.16626522],
        [-0.13633697, -0.10083474],
        [-0.3676528 , -0.12974691]]),
 array([1.73650684, 0.31403256]),
 array([[-0.74885926, -0.66272906],
        [-0.66272906,  0.74885926]]))

rr = np.linalg.matrix_rank(X)
print('rank of X - '), rr

rank of X -

(None, 2)

Properties:

Where $U$ is constructed via a full SVD, $U^T U = I_{p\times p}$ and $U U^T = I_{m \times m}$
Where $\hat U$ is constructed via a reduced SVD, although $\hat U^T \hat U = I_{p\times p}$ it happens that $\hat U \hat U^T \neq I_{m \times m}$

We illustrate these properties for our example with the following code cells.

UTU = U.T@U
UUT = U@U.T
print('UUT, UTU = '), UUT, UTU 

UUT, UTU =

(None,
 array([[ 1.00000000e+00, -7.21936265e-17, -1.54572027e-16,
         -3.31764657e-17, -1.38796421e-16],
        [-7.21936265e-17,  1.00000000e+00,  2.71193549e-17,
         -3.40281676e-17,  1.11277046e-17],
        [-1.54572027e-16,  2.71193549e-17,  1.00000000e+00,
          4.29425949e-18,  2.47458879e-17],
        [-3.31764657e-17, -3.40281676e-17,  4.29425949e-18,
          1.00000000e+00, -1.23447957e-17],
        [-1.38796421e-16,  1.11277046e-17,  2.47458879e-17,
         -1.23447957e-17,  1.00000000e+00]]),
 array([[ 1.00000000e+00,  6.91569287e-17, -3.92036831e-16,
         -5.66121804e-17, -1.51775901e-16],
        [ 6.91569287e-17,  1.00000000e+00,  4.87605731e-17,
         -4.16965272e-17, -8.94647853e-17],
        [-3.92036831e-16,  4.87605731e-17,  1.00000000e+00,
         -2.05891352e-17, -8.66260968e-17],
        [-5.66121804e-17, -4.16965272e-17, -2.05891352e-17,
          1.00000000e+00, -2.69818135e-17],
        [-1.51775901e-16, -8.94647853e-17, -8.66260968e-17,
         -2.69818135e-17,  1.00000000e+00]]))

UhatUhatT = Uhat@Uhat.T
UhatTUhat = Uhat.T@Uhat
print('UhatUhatT, UhatTUhat= '), UhatUhatT, UhatTUhat

UhatUhatT, UhatTUhat=

(None,
 array([[ 0.51764759, -0.199257  ,  0.43396176,  0.02929958,  0.14423569],
        [-0.199257  ,  0.85744142,  0.08341717,  0.13417202,  0.23994108],
        [ 0.43396176,  0.08341717,  0.44415273,  0.07122313,  0.21570144],
        [ 0.02929958,  0.13417202,  0.07122313,  0.02875541,  0.06320766],
        [ 0.14423569,  0.23994108,  0.21570144,  0.06320766,  0.15200284]]),
 array([[1.00000000e+00, 6.91569287e-17],
        [6.91569287e-17, 1.00000000e+00]]))

Remarks:

The cells above illustrate application of the fullmatrices=True and full-matrices=False options. Using full-matrices=False returns a reduced singular value decomposition.

This option implements an optimal reduced rank approximation of a matrix, in the sense of minimizing the Frobenius norm of the discrepancy between the approximating matrix and the matrix being approximated.

Optimality in this sense is established in the celebrated Eckart–Young theorem. See https://en.wikipedia.org/wiki/Low-rank_approximation.

When we study Dynamic Mode Decompositions below, it will be important for us to remember the preceding properties of full and reduced SVD’s in such tall-skinny cases.

Now let’s turn to a short-fat case.

To illustrate this case, we’ll set $m = 2 < 5 = n $

import numpy as np
X = np.random.rand(2,5)
U, S, V = np.linalg.svd(X,full_matrices=True)  # full SVD
Uhat, Shat, Vhat = np.linalg.svd(X,full_matrices=False) # economy SVD
print('U, S, V ='), U, S, V

U, S, V =

(None,
 array([[-0.65304737, -0.75731706],
        [-0.75731706,  0.65304737]]),
 array([1.72915162, 0.61969749]),
 array([[-0.59925033, -0.02198324, -0.43071446, -0.34202447, -0.58130895],
        [-0.20938284,  0.00244402,  0.26574177,  0.81917651, -0.46312418],
        [-0.302653  , -0.40764758,  0.78360476, -0.35530191, -0.04414456],
        [-0.0166243 , -0.88787434, -0.35804252,  0.22001626,  0.18655074],
        [-0.71076203,  0.21216352, -0.04044126,  0.19319032,  0.64097268]]))

print('Uhat, Shat, Vhat = '), Uhat, Shat, Vhat

Uhat, Shat, Vhat =

(None,
 array([[-0.65304737, -0.75731706],
        [-0.75731706,  0.65304737]]),
 array([1.72915162, 0.61969749]),
 array([[-0.59925033, -0.02198324, -0.43071446, -0.34202447, -0.58130895],
        [-0.20938284,  0.00244402,  0.26574177,  0.81917651, -0.46312418]]))

rr = np.linalg.matrix_rank(X)
print('rank X = '), rr

rank X =

(None, 2)

7.5. Digression: Polar Decomposition¶

A singular value decomposition (SVD) of $X$ is related to a polar decomposition of $X$

\[ X = SQ \]

where

\[ \begin{aligned} S & = U\Sigma U^T \cr Q & = U V^T \end{aligned} \]

and $S$ is evidently a symmetric matrix and $Q$ is an orthogonal matrix.

7.6. Principal Components Analysis (PCA)¶

Let’s begin with a case in which $n >> m$, so that we have many more individuals $n$ than attributes $m$.

The matrix $X$ is short and fat in an $n >> m$ case as opposed to a tall and skinny case with $m > > n $ to be discussed later.

We regard $X$ as an $m \times n$ matrix of data:

\[ X = \begin{bmatrix} X_1 \mid X_2 \mid \cdots \mid X_n\end{bmatrix} \]

where for $j = 1, \ldots, n$ the column vector $X_j = \begin{bmatrix}X_{1j}\\X_{2j}\\\vdots\\X_{mj}\end{bmatrix}$ is a vector of observations on variables $\begin{bmatrix}x_1\\x_2\\\vdots\\x_m\end{bmatrix}$.

In a time series setting, we would think of columns $j$ as indexing different times at which random variables are observed, while rows index different random variables.

In a cross section setting, we would think of columns $j$ as indexing different individuals for which random variables are observed, while rows index different attributes.

The number of positive singular values equals the rank of matrix $X$.

Arrange the singular values in decreasing order.

Arrange the positive singular values on the main diagonal of the matrix $\Sigma$ of into a vector $\sigma_R$.

Set all other entries of $\Sigma$ to zero.

7.7. Relationship of PCA to SVD¶

To relate a SVD to a PCA (principal component analysis) of data set $X$, first construct the SVD of the data matrix $X$:

(7.1)¶\[ X = U \Sigma V^T = \sigma_1 U_1 V_1^T + \sigma_2 U_2 V_2^T + \cdots + \sigma_p U_p V_p^T \]

where

\[ U=\begin{bmatrix}U_1|U_2|\ldots|U_m\end{bmatrix} \]

\[\begin{split} V^T = \begin{bmatrix}V_1^T\\V_2^T\\\ldots\\V_n^T\end{bmatrix} \end{split}\]

In equation (7.1), each of the $m \times n$ matrices $U_{j}V_{j}^T$ is evidently of rank $1$.

Thus, we have

(7.2)¶\[\begin{split} X = \sigma_1 \begin{pmatrix}U_{11}V_{1}^T\\U_{21}V_{1}^T\\\cdots\\U_{m1}V_{1}^T\\\end{pmatrix} + \sigma_2\begin{pmatrix}U_{12}V_{2}^T\\U_{22}V_{2}^T\\\cdots\\U_{m2}V_{2}^T\\\end{pmatrix}+\ldots + \sigma_p\begin{pmatrix}U_{1p}V_{p}^T\\U_{2p}V_{p}^T\\\cdots\\U_{mp}V_{p}^T\\\end{pmatrix} \end{split}\]

Here is how we would interpret the objects in the matrix equation (7.2) in a time series context:

$ \textrm{for each} \ k=1, \ldots, n $, the object $\lbrace V_{kj} \rbrace_{j=1}^n$ is a time series for the $k$th principal component
$U_j = \begin{bmatrix}U_{1k}\\U_{2k}\\\ldots\\U_{mk}\end{bmatrix} \ k=1, \ldots, m$ is a vector of loadings of variables $X_i$ on the $k$th principal component, $i=1, \ldots, m$
$\sigma_k $ for each $k=1, \ldots, p$ is the strength of $k$th principal component, where strength means contribution to the overall covariance of $X$.

7.8. PCA with Eigenvalues and Eigenvectors¶

We now use an eigen decomposition of a sample covariance matrix to do PCA.

Let $X_{m \times n}$ be our $m \times n$ data matrix.

Let’s assume that sample means of all variables are zero.

We can assure this by pre-processing the data by subtracting sample means.

Define a sample covariance matrix $\Omega$ as

\[ \Omega = XX^T \]

Then use an eigen decomposition to represent $\Omega$ as follows:

\[ \Omega =P\Lambda P^T \]

Here

$P$ is $m×m$ matrix of eigenvectors of $\Omega$
$\Lambda$ is a diagonal matrix of eigenvalues of $\Omega$

We can then represent $X$ as

\[ X=P\epsilon \]

where

\[ \epsilon = P^{-1} X \]

and

\[ \epsilon\epsilon^T=\Lambda . \]

We can verify that

(7.3)¶\[ XX^T=P\Lambda P^T . \]

It follows that we can represent the data matrix $X$ as

\[\begin{equation*} X=\begin{bmatrix}X_1|X_2|\ldots|X_m\end{bmatrix} =\begin{bmatrix}P_1|P_2|\ldots|P_m\end{bmatrix} \begin{bmatrix}\epsilon_1\\\epsilon_2\\\ldots\\\epsilon_m\end{bmatrix} = P_1\epsilon_1+P_2\epsilon_2+\ldots+P_m\epsilon_m \end{equation*}\]

To reconcile the preceding representation with the PCA that we had obtained earlier through the SVD, we first note that $\epsilon_j^2=\lambda_j\equiv\sigma^2_j$.

Now define $\tilde{\epsilon_j} = \frac{\epsilon_j}{\sqrt{\lambda_j}}$, which implies that $\tilde{\epsilon}_j\tilde{\epsilon}_j^T=1$.

Therefore

\[\begin{split} \begin{aligned} X&=\sqrt{\lambda_1}P_1\tilde{\epsilon_1}+\sqrt{\lambda_2}P_2\tilde{\epsilon_2}+\ldots+\sqrt{\lambda_m}P_m\tilde{\epsilon_m}\\ &=\sigma_1P_1\tilde{\epsilon_2}+\sigma_2P_2\tilde{\epsilon_2}+\ldots+\sigma_mP_m\tilde{\epsilon_m} , \end{aligned} \end{split}\]

which agrees with

\[ X=\sigma_1U_1{V_1}^{T}+\sigma_2 U_2{V_2}^{T}+\ldots+\sigma_{r} U_{r}{V_{r}}^{T} \]

provided that we set

$U_j=P_j$ (a vector of loadings of variables on principal component $j$)
${V_k}^{T}=\tilde{\epsilon_k}$ (the $k$th principal component)

Because there are alternative algorithms for computing $P$ and $U$ for given a data matrix $X$, depending on algorithms used, we might have sign differences or different orders between eigenvectors.

We can resolve such ambiguities about $U$ and $P$ by

sorting eigenvalues and singular values in descending order
imposing positive diagonals on $P$ and $U$ and adjusting signs in $V^T$ accordingly

7.9. Connections¶

To pull things together, it is useful to assemble and compare some formulas presented above.

First, consider an SVD of an $m \times n$ matrix:

\[ X = U\Sigma V^T \]

Compute:

(7.4)¶\[ \begin{aligned} XX^T&=U\Sigma V^TV\Sigma^T U^T\cr &\equiv U\Sigma\Sigma^TU^T\cr &\equiv U\Lambda U^T \end{aligned} \]

Compare representation (7.4) with equation (7.3) above.

Evidently, $U$ in the SVD is the matrix $P$ of eigenvectors of $XX^T$ and $\Sigma \Sigma^T$ is the matrix $\Lambda$ of eigenvalues.

Second, let’s compute

\[\begin{split} \begin{aligned} X^TX &=V\Sigma^T U^TU\Sigma V^T\\ &=V\Sigma^T{\Sigma}V^T \end{aligned} \end{split}\]

Thus, the matrix $V$ in the SVD is the matrix of eigenvectors of $X^TX$

Summarizing and fitting things together, we have the eigen decomposition of the sample covariance matrix

\[ X X^T = P \Lambda P^T \]

where $P$ is an orthogonal matrix.

Further, from the SVD of $X$, we know that

\[ X X^T = U \Sigma \Sigma^T U^T \]

where $U$ is an orthonal matrix.

Thus, $P = U$ and we have the representation of $X$

\[ X = P \epsilon = U \Sigma V^T \]

It follows that

\[ U^T X = \Sigma V^T = \epsilon \]

Note that the preceding implies that

\[ \epsilon \epsilon^T = \Sigma V^T V \Sigma^T = \Sigma \Sigma^T = \Lambda , \]

so that everything fits together.

Below we define a class DecomAnalysis that wraps PCA and SVD for a given a data matrix X.

class DecomAnalysis:
    """
    A class for conducting PCA and SVD.
    """

    def __init__(self, X, n_component=None):

        self.X = X

        self.Ω = (X @ X.T)

        self.m, self.n = X.shape
        self.r = LA.matrix_rank(X)

        if n_component:
            self.n_component = n_component
        else:
            self.n_component = self.m

    def pca(self):

        𝜆, P = LA.eigh(self.Ω)    # columns of P are eigenvectors

        ind = sorted(range(𝜆.size), key=lambda x: 𝜆[x], reverse=True)

        # sort by eigenvalues
        self.𝜆 = 𝜆[ind]
        P = P[:, ind]
        self.P = P @ diag_sign(P)

        self.Λ = np.diag(self.𝜆)

        self.explained_ratio_pca = np.cumsum(self.𝜆) / self.𝜆.sum()

        # compute the N by T matrix of principal components 
        self.𝜖 = self.P.T @ self.X

        P = self.P[:, :self.n_component]
        𝜖 = self.𝜖[:self.n_component, :]

        # transform data
        self.X_pca = P @ 𝜖

    def svd(self):

        U, 𝜎, VT = LA.svd(self.X)

        ind = sorted(range(𝜎.size), key=lambda x: 𝜎[x], reverse=True)

        # sort by eigenvalues
        d = min(self.m, self.n)

        self.𝜎 = 𝜎[ind]
        U = U[:, ind]
        D = diag_sign(U)
        self.U = U @ D
        VT[:d, :] = D @ VT[ind, :]
        self.VT = VT

        self.Σ = np.zeros((self.m, self.n))
        self.Σ[:d, :d] = np.diag(self.𝜎)

        𝜎_sq = self.𝜎 ** 2
        self.explained_ratio_svd = np.cumsum(𝜎_sq) / 𝜎_sq.sum()

        # slicing matrices by the number of components to use
        U = self.U[:, :self.n_component]
        Σ = self.Σ[:self.n_component, :self.n_component]
        VT = self.VT[:self.n_component, :]

        # transform data
        self.X_svd = U @ Σ @ VT

    def fit(self, n_component):

        # pca
        P = self.P[:, :n_component]
        𝜖 = self.𝜖[:n_component, :]

        # transform data
        self.X_pca = P @ 𝜖

        # svd
        U = self.U[:, :n_component]
        Σ = self.Σ[:n_component, :n_component]
        VT = self.VT[:n_component, :]

        # transform data
        self.X_svd = U @ Σ @ VT

def diag_sign(A):
    "Compute the signs of the diagonal of matrix A"

    D = np.diag(np.sign(np.diag(A)))

    return D

We also define a function that prints out information so that we can compare decompositions obtained by different algorithms.

def compare_pca_svd(da):
    """
    Compare the outcomes of PCA and SVD.
    """

    da.pca()
    da.svd()

    print('Eigenvalues and Singular values\n')
    print(f'λ = {da.λ}\n')
    print(f'σ^2 = {da.σ**2}\n')
    print('\n')

    # loading matrices
    fig, axs = plt.subplots(1, 2, figsize=(14, 5))
    plt.suptitle('loadings')
    axs[0].plot(da.P.T)
    axs[0].set_title('P')
    axs[0].set_xlabel('m')
    axs[1].plot(da.U.T)
    axs[1].set_title('U')
    axs[1].set_xlabel('m')
    plt.show()

    # principal components
    fig, axs = plt.subplots(1, 2, figsize=(14, 5))
    plt.suptitle('principal components')
    axs[0].plot(da.ε.T)
    axs[0].set_title('ε')
    axs[0].set_xlabel('n')
    axs[1].plot(da.VT[:da.r, :].T * np.sqrt(da.λ))
    axs[1].set_title('$V^T*\sqrt{\lambda}$')
    axs[1].set_xlabel('n')
    plt.show()

For an example PCA applied to analyzing the structure of intelligence tests see this lecture Multivariable Normal Distribution.

Look at parts of that lecture that describe and illustrate the classic factor analysis model.

7.10. Dynamic Mode Decomposition (DMD)¶

We turn to the tall and skinny case associated with Dynamic Mode Decomposition, the case in which $ m >>n $.

Here an $ m \times n $ data matrix $ \tilde X $ contains many more attributes $ m $ than individuals $ n $.

This

Dynamic mode decomposition was introduced by [Sch10],

You can read more about Dynamic Mode Decomposition here [KBBWP16] and here [BK19] (section 7.2).

We want to fit a first-order vector autoregression

(7.5)¶\[ X_{t+1} = A X_t + C \epsilon_{t+1} \]

where $\epsilon_{t+1}$ is the time $t+1$ instance of an i.i.d. $m \times 1$ random vector with mean vector zero and identity covariance matrix and

where the $ m \times 1 $ vector $ X_t $ is

(7.6)¶\[ X_t = \begin{bmatrix} X_{1,t} & X_{2,t} & \cdots & X_{m,t} \end{bmatrix}^T \]

and where $ T $ again denotes complex transposition and $ X_{i,t} $ is an observation on variable $ i $ at time $ t $.

We want to fit equation (7.5).

Our data are organized in an $ m \times (n+1) $ matrix $ \tilde X $

\[ \tilde X = \begin{bmatrix} X_1 \mid X_2 \mid \cdots \mid X_n \mid X_{n+1} \end{bmatrix} \]

where for $ t = 1, \ldots, n +1 $, the $ m \times 1 $ vector $ X_t $ is given by (7.6).

Thus, we want to estimate a system (7.5) that consists of $ m $ least squares regressions of everything on one lagged value of everything.

The $i$’th equation of (7.5) is a regression of $X_{i,t+1}$ on the vector $X_t$.

We proceed as follows.

From $ \tilde X $, we form two $m \times n$ matrices

\[ X = \begin{bmatrix} X_1 \mid X_2 \mid \cdots \mid X_{n}\end{bmatrix} \]

and

\[ X' = \begin{bmatrix} X_2 \mid X_3 \mid \cdots \mid X_{n+1}\end{bmatrix} \]

Here $ ' $ does not indicate matrix transposition but instead is part of the name of the matrix $ X' $.

In forming $ X $ and $ X' $, we have in each case dropped a column from $ \tilde X $, the last column in the case of $ X $, and the first column in the case of $ X' $.

Evidently, $ X $ and $ X' $ are both $ m \times n $ matrices.

We denote the rank of $ X $ as $ p \leq \min(m, n) $.

Two possible cases are

$ n > > m$, so that we have many more time series observations $n$ than variables $m$
$m > > n$, so that we have many more variables $m $ than time series observations $n$

At a general level that includes both of these special cases, a common formula describes the least squares estimator $\hat A$ of $A$ for both cases.

But some important details differ.

The common formula is

(7.7)¶\[ \hat A = X' X^+ \]

where $X^+$ is the pseudo-inverse of $X$.

Applicable formulas for the pseudo-inverse differ for our two cases.

Short-Fat Case:

When $ n > > m$, so that we have many more time series observations $n$ than variables $m$ and when $X$ has linearly independent rows, $X X^T$ has an inverse and the pseudo-inverse $X^+$ is

\[ X^+ = X^T (X X^T)^{-1} \]

Here $X^+$ is a right-inverse that verifies $ X X^+ = I_{m \times m}$.

In this case, our formula (7.7) for the least-squares estimator of the population matrix of regression coefficients $A$ becomes

(7.8)¶\[ \hat A = X' X^T (X X^T)^{-1} \]

This formula for least-squares regression coefficients widely used in econometrics.

For example, it is used to estimate vector autorgressions.

The right side of formula (7.8) is proportional to the empirical cross second moment matrix of $X_{t+1}$ and $X_t$ times the inverse of the second moment matrix of $X_t$.

Tall-Skinny Case:

When $m > > n$, so that we have many more attributes $m $ than time series observations $n$ and when $X$ has linearly independent columns, $X^T X$ has an inverse and the pseudo-inverse $X^+$ is

\[ X^+ = (X^T X)^{-1} X^T \]

Here $X^+$ is a left-inverse that verifies $X^+ X = I_{n \times n}$.

In this case, our formula (7.7) for a least-squares estimator of $A$ becomes

(7.9)¶\[ \hat A = X' (X^T X)^{-1} X^T \]

Please compare formulas (7.8) and (7.9) for $\hat A$.

Here we are interested in formula (7.9).

The $ i $th row of $ \hat A $ is an $ m \times 1 $ vector of regression coefficients of $ X_{i,t+1} $ on $ X_{j,t}, j = 1, \ldots, m $.

If we use formula (7.9) to calculate $\hat A X$ we find that

\[ \hat A X = X' \]

so that the regression equation fits perfectly.

This is the usual outcome in an underdetermined least-squares model.

To reiterate, in our tall-skinny case in which we have a number $n$ of observations that is small relative to the number $m$ of attributes that appear in the vector $X_t$, we want to fit equation (7.5).

To offer ideas about how we can efficiently calculate the pseudo-inverse $X^+$, as our estimator $\hat A$ of $A$ we form an $m \times m$ matrix that solves the least-squares best-fit problem

(7.10)¶\[ \hat A = \textrm{argmin}_{\check A} || X' - \check A X ||_F \]

where $|| \cdot ||_F$ denotes the Frobenius (or Euclidean) norm of a matrix.

The Frobenius norm is defined as

\[ ||A||_F = \sqrt{ \sum_{i=1}^m \sum_{j=1}^m |A_{ij}|^2 } \]

The minimizer of the right side of equation (7.10) is

(7.11)¶\[ \hat A = X' X^{+} \]

where the (possibly huge) $ n \times m $ matrix $ X^{+} = (X^T X)^{-1} X^T$ is again a pseudo-inverse of $ X $.

For some situations that we are interested in, $X^T X $ can be close to singular, a situation that can make some numerical algorithms be error-prone.

To acknowledge that possibility, we’ll use efficient algorithms for computing and for constructing reduced rank approximations of $\hat A$ in formula (7.9).

The $ i $th row of $ \hat A $ is an $ m \times 1 $ vector of regression coefficients of $ X_{i,t+1} $ on $ X_{j,t}, j = 1, \ldots, m $.

An efficient way to compute the pseudo-inverse $X^+$ is to start with a singular value decomposition

(7.12)¶\[ X = U \Sigma V^T \]

where we remind ourselves that for a reduced SVD, $X$ is an $m \times n$ matrix of data, $U$ is an $m \times p$ matrix, $\Sigma$ is a $p \times p$ matrix, and $V is an $n \times p$ matrix.

We can efficiently construct the pertinent pseudo-inverse $X^+$ by recognizing the following string of equalities.

(7.13)¶\[\begin{split} \begin{aligned} X^{+} & = (X^T X)^{-1} X^T \\ & = (V \Sigma U^T U \Sigma V^T)^{-1} V \Sigma U^T \\ & = (V \Sigma \Sigma V^T)^{-1} V \Sigma U^T \\ & = V \Sigma^{-1} \Sigma^{-1} V^T V \Sigma U^T \\ & = V \Sigma^{-1} U^T \end{aligned} \end{split}\]

(Since we are in the $m > > n$ case in which $V^T V = I_{p \times p}$ in a reduced SVD, we can use the preceding string of equalities for a reduced SVD as well as for a full SVD.)

Thus, we shall construct a pseudo-inverse $ X^+ $ of $ X $ by using a singular value decomposition of $X$ in equation (7.12) to compute

(7.14)¶\[ X^{+} = V \Sigma^{-1} U^T \]

where the matrix $ \Sigma^{-1} $ is constructed by replacing each non-zero element of $ \Sigma $ with $ \sigma_j^{-1} $.

We can use formula (7.14) together with formula (7.11) to compute the matrix $ \hat A $ of regression coefficients.

Thus, our estimator $\hat A = X' X^+$ of the $m \times m$ matrix of coefficients $A$ is

(7.15)¶\[ \hat A = X' V \Sigma^{-1} U^T \]

We’ll eventually use dynamic mode decomposition to compute a rank $ r $ approximation to $ \hat A $, where $ r < p $.

Remark: In our Python code, we’ll sometimes use a reduced SVD.

Next, we describe alternative representations of our first-order linear dynamic system.

Guide to three representations: In practice, we’ll be interested in Representation 3. We present the first 2 in order to set the stage for some intermediate steps that might help us understand what is under the hood of Representation 3. In applications, we’ll use only a small subset of the DMD to approximate dynamics. To to that, we’ll want to be using the reduced SVD’s affiliated with representation 3, not the full SVD’s affiliated with Representations 1 and 2.

7.11. Representation 1¶

In this representation, we shall use a full SVD of $X$.

We use the $m$ columns of $U$, and thus the $m$ rows of $U^T$, to define a $m \times 1$ vector $\tilde b_t$ as

(7.16)¶\[ \tilde b_t = U^T X_t . \]

The original data $X_t$ can be represented as

(7.17)¶\[ X_t = U \tilde b_t \]

(Here we use $b$ to remind ourselves that we are creating a basis vector.)

Since we are now using a full SVD, $U U^T = I_{m \times m}$.

So it follows from equation (7.16) that we can reconstruct $X_t$ from $\tilde b_t$.

In particular,

Equation (7.16) serves as an encoder that rotates the $m \times 1$ vector $X_t$ to become an $m \times 1$ vector $\tilde b_t$
Equation (7.17) serves as a decoder that reconstructs the $m \times 1$ vector $X_t$ by rotating the $m \times 1$ vector $\tilde b_t$

Define a transition matrix for an $m \times 1$ basis vector $\tilde b_t$ by

(7.18)¶\[ \tilde A = U^T \hat A U \]

We can recover $\hat A$ from

\[ \hat A = U \tilde A U^T \]

Dynamics of the $m \times 1$ basis vector $\tilde b_t$ are governed by

\[ \tilde b_{t+1} = \tilde A \tilde b_t \]

To construct forecasts $\overline X_t$ of future values of $X_t$ conditional on $X_1$, we can apply decoders (i.e., rotators) to both sides of this equation and deduce

\[ \overline X_{t+1} = U \tilde A^t U^T X_1 \]

where we use $\overline X_{t+1}, t \geq 1 $ to denote a forecast.

7.12. Representation 2¶

This representation is related to one originally proposed by [Sch10].

It can be regarded as an intermediate step to a related representation 3 to be presented later

As with Representation 1, we continue to

use a full SVD and not a reduced SVD

As we observed and illustrated earlier in this lecture

(a) for a full SVD $U U^T = I_{m \times m} $ and $U^T U = I_{p \times p}$ are both identity matrices
(b) for a reduced SVD of $X$, $U^T U $ is not an identity matrix.

As we shall see later, a full SVD is too confining for what we ultimately want to do, namely, situations in which $U^T U$ is not an identity matrix because we use a reduced SVD of $X$.

But for now, let’s proceed under the assumption that we are using a full SVD so that both of the preceding two requirements (a) and (b) are satisfied.

Form an eigendecomposition of the $m \times m$ matrix $\tilde A = U^T \hat A U$ defined in equation (7.18):

(7.19)¶\[ \tilde A = W \Lambda W^{-1} \]

where $\Lambda$ is a diagonal matrix of eigenvalues and $W$ is an $m \times m$ matrix whose columns are eigenvectors corresponding to rows (eigenvalues) in $\Lambda$.

When $U U^T = I_{m \times m}$, as is true with a full SVD of $X$, it follows that

(7.20)¶\[ \hat A = U \tilde A U^T = U W \Lambda W^{-1} U^T \]

According to equation (7.20), the diagonal matrix $\Lambda$ contains eigenvalues of $\hat A$ and corresponding eigenvectors of $\hat A$ are columns of the matrix $UW$.

It follows that the systematic (i.e., not random) parts of the $X_t$ dynamics captured by our first-order vector autoregressions are described by

\[ X_{t+1} = U W \Lambda W^{-1} U^T X_t \]

Multiplying both sides of the above equation by $W^{-1} U^T$ gives

\[ W^{-1} U^T X_{t+1} = \Lambda W^{-1} U^T X_t \]

or

\[ \hat b_{t+1} = \Lambda \hat b_t \]

where our encoder is now

\[ \hat b_t = W^{-1} U^T X_t \]

and our decoder is

\[ X_t = U W \hat b_t \]

We can use this representation to construct a predictor $\overline X_{t+1}$ of $X_{t+1}$ conditional on $X_1$ via:

(7.21)¶\[ \overline X_{t+1} = U W \Lambda^t W^{-1} U^T X_1 \]

In effect, [Sch10] defined an $m \times m$ matrix $\Phi_s$ as

(7.22)¶\[ \Phi_s = UW \]

and a generalized inverse

(7.23)¶\[ \Phi_s^+ = W^{-1}U^T \]

[Sch10] then represented equation (7.21) as

(7.24)¶\[ \overline X_{t+1} = \Phi_s \Lambda^t \Phi_s^+ X_1 \]

Components of the basis vector $ \hat b_t = W^{-1} U^T X_t \equiv \Phi_s^+ X_t$ are often called DMD modes, or sometimes also DMD projected modes.

To understand why they are called projected modes, notice that

\[ \Phi_s^+ = ( \Phi_s^T \Phi_s)^{-1} \Phi_s^T \]

so that the $m \times p$ matrix

\[ \hat b = \Phi_s^+ X \]

is a matrix of regression coefficients of the $m \times n$ matrix $X$ on the $m \times p$ matrix $\Phi_s$.

We’ll say more about this interpretation in a related context when we discuss representation 3.

We turn next to an alternative representation suggested by Tu et al. [TRL+14].

It is more appropriate to use this alternative representation when, as in practice is typically the case, we use a reduced SVD.

7.13. Representation 3¶

Departing from the procedures used to construct Representations 1 and 2, each of which deployed a full SVD, we now use a reduced SVD.

Again, we let $p \leq \textrm{min}(m,n)$ be the rank of $X$.

Construct a reduced SVD

\[ X = \tilde U \tilde \Sigma \tilde V^T, \]

where now $\tilde U$ is $m \times p$, $\tilde \Sigma$ is $ p \times p$, and $\tilde V^T$ is $p \times n$.

Our minimum-norm least-squares estimator approximator of $A$ now has representation

(7.25)¶\[ \hat A = X' \tilde V \tilde \Sigma^{-1} \tilde U^T \]

Paralleling a step in Representation 1, define a transition matrix for a rotated $p \times 1$ state $\tilde b_t$ by

(7.26)¶\[ \tilde A =\tilde U^T \hat A \tilde U \]

Interpretation as projection coefficients

[BK22] remark that $\tilde A$ can be interpreted in terms of a projection of $\hat A$ onto the $p$ modes in $\tilde U$.

To verify this, first note that, because $ \tilde U^T \tilde U = I$, it follows that

(7.27)¶\[ \tilde A = \tilde U^T \hat A \tilde U = \tilde U^T X' \tilde V \tilde \Sigma^{-1} \tilde U^T \tilde U = \tilde U^T X' \tilde V \tilde \Sigma^{-1} \tilde U^T \]

Next, we’ll just compute the regression coefficients in a projection of $\hat A$ on $\tilde U$ using the standard least-square formula

\[ (\tilde U^T \tilde U)^{-1} \tilde U^T \hat A = (\tilde U^T \tilde U)^{-1} \tilde U^T X' \tilde V \tilde \Sigma^{-1} \tilde U^T = \tilde U^T X' \tilde V \tilde \Sigma^{-1} \tilde U^T = \tilde A . \]

Note that because we are now working with a reduced SVD, $\tilde U \tilde U^T \neq I$.

Consequently,

\[ \hat A \neq \tilde U \tilde A \tilde U^T, \]

and we can’t simply recover $\hat A$ from $\tilde A$ and $\tilde U$.

Nevertheless, we hope for the best and proceed to construct an eigendecomposition of the $p \times p$ matrix $\tilde A$:

(7.28)¶\[ \tilde A = \tilde W \Lambda \tilde W^{-1} . \]

Mimicking our procedure in Representation 2, we cross our fingers and compute an $m \times p$ matrix

(7.29)¶\[ \tilde \Phi_s = \tilde U \tilde W \]

that corresponds to (7.22) for a full SVD.

At this point, where $\hat A$ is given by formula (7.25) it is interesting to compute $\hat A \tilde \Phi_s$:

\[\begin{split} \begin{aligned} \hat A \tilde \Phi_s & = (X' \tilde V \tilde \Sigma^{-1} \tilde U^T) (\tilde U \tilde W) \\ & = X' \tilde V \tilde \Sigma^{-1} \tilde W \\ & \neq (\tilde U \tilde W) \Lambda \\ & = \tilde \Phi_s \Lambda \end{aligned} \end{split}\]

That $ \hat A \tilde \Phi_s \neq \tilde \Phi_s \Lambda $ means, that unlike the corresponding situation in Representation 2, columns of $\tilde \Phi_s = \tilde U \tilde W$ are not eigenvectors of $\hat A$ corresponding to eigenvalues on the diagonal of matix $\Lambda$.

But in a quest for eigenvectors of $\hat A$ that we can compute with a reduced SVD, let’s define the $m \times p$ matrix $\Phi$ as

(7.30)¶\[ \Phi \equiv \hat A \tilde \Phi_s = X' \tilde V \tilde \Sigma^{-1} \tilde W \]

It turns out that columns of $\Phi$ are eigenvectors of $\hat A$.

This is a consequence of a result established by Tu et al. [TRL+14], which we now present.

Proposition The $p$ columns of $\Phi$ are eigenvectors of $\hat A$.

Proof: From formula (7.30) we have

\[ \begin{aligned} \hat A \Phi & = (X' \tilde V \tilde \Sigma^{-1} \tilde U^T) (X' \tilde V \Sigma^{-1} \tilde W) \cr & = X' \tilde V \tilde \Sigma^{-1} \tilde A \tilde W \cr & = X' \tilde V \tilde \Sigma^{-1}\tilde W \Lambda \cr & = \Phi \Lambda \end{aligned} \]

Thus, we have deduced that

(7.31)¶\[ \hat A \Phi = \Phi \Lambda \]

Let $\phi_i$ be the $i$th column of $\Phi$ and $\lambda_i$ be the corresponding $i$ eigenvalue of $\tilde A$ from decomposition (7.28).

Writing out the $m \times 1$ vectors on both sides of equation (7.31) and equating them gives

\[ \hat A \phi_i = \lambda_i \phi_i . \]

This equation confirms that $\phi_i$ is an eigenvector of $\hat A$ that corresponds to eigenvalue $\lambda_i$ of both $\tilde A$ and $\hat A$.

This concludes the proof.

Also see [BK22] (p. 238)

7.13.1. Decoder of $X$ as a linear projection¶

From eigendecomposition (7.31) we can represent $\hat A$ as

(7.32)¶\[ \hat A = \Phi \Lambda \Phi^+ . \]

From formula (7.32) we can deduce the reduced dimension dynamics

\[ \check b_{t+1} = \Lambda \check b_t \]

where

(7.33)¶\[ \check b_t = \Phi^+ X_t \]

Since the $m \times p$ matrix $\Phi$ has $p$ linearly independent columns, the generalized inverse of $\Phi$ is

\[ \Phi^{+} = (\Phi^T \Phi)^{-1} \Phi^T \]

and so

(7.34)¶\[ \check b = (\Phi^T \Phi)^{-1} \Phi^T X \]

The $p \times n$ matrix $\check b$ is recognizable as a matrix of least squares regression coefficients of the $m \times n$ matrix $X$ on the $m \times p$ matrix $\Phi$ and consequently

(7.35)¶\[ \check X = \Phi \check b \]

is an $m \times n$ matrix of least squares projections of $X$ on $\Phi$.

By virtue of least-squares projection theory discussed here https://python-advanced.quantecon.org/orth_proj.html, we can represent $X$ as the sum of the projection $\check X$ of $X$ on $\Phi$ plus a matrix of errors.

To verify this, note that the least squares projection $\check X$ is related to $X$ by

\[ X = \check X + \epsilon \]

or

(7.36)¶\[ X = \Phi \check b + \epsilon \]

where $\epsilon$ is an $m \times n$ matrix of least squares errors satisfying the least squares orthogonality conditions $\epsilon^T \Phi =0 $ or

(7.37)¶\[ (X - \Phi \check b)^T \Phi = 0_{m \times p} \]

Rearranging the orthogonality conditions (7.37) gives $X^T \Phi = \check b \Phi^T \Phi$, which implies formula (7.34).

7.13.2. A useful approximation¶

There is a useful way to approximate the $p \times 1$ vector $\check b_t$ instead of using formula (7.33).

In particular, the following argument adapted from [BK22] (page 240) provides a computationally efficient way to approximate $\check b_t$.

For convenience, we’ll do this first for time $t=1$.

For $t=1$, from equation (7.36) we have

(7.38)¶\[ \check X_1 = \Phi \check b_1 \]

where $\check b_1$ is a $p \times 1$ vector.

Recall from representation 1 above that $X_1 = U \tilde b_1$, where $\tilde b_1$ is a time $1$ basis vector for representation 1 and $U$ is from a full SVD of $X$.

It then follows from equation (7.36) that

\[ U \tilde b_1 = X' \tilde V \tilde \Sigma^{-1} \tilde W \check b_1 + \epsilon_1 \]

where $\epsilon_1$ is a least-squares error vector from equation (7.36).

It follows that

\[ \tilde b_1 = U^T X' V \tilde \Sigma^{-1} \tilde W \check b_1 + U^T \epsilon_1 \]

Replacing the error term $U^T \epsilon_1$ by zero, and replacing $U$ from a full SVD of $X$ with $\tilde U$ from a reduced SVD, we obtain an approximation $\hat b_1$ to $\tilde b_1$:

\[ \hat b_1 = \tilde U^T X' \tilde V \tilde \Sigma^{-1} \tilde W \check b_1 \]

Recall that from equation (7.27), $ \tilde A = \tilde U^T X' \tilde V \tilde \Sigma^{-1}$.

It then follows that

\[ \hat b_1 = \tilde A \tilde W \check b_1 \]

and therefore, by the eigendecomposition (7.28) of $\tilde A$, we have

\[ \hat b_1 = \tilde W \Lambda \check b_1 \]

Consequently,

\[ \hat b_1 = ( \tilde W \Lambda)^{-1} \tilde b_1 \]

or

(7.39)¶\[ \hat b_1 = ( \tilde W \Lambda)^{-1} \tilde U^T X_1 , \]

which is computationally efficient approximation to the following instance of equation (7.33) for the initial vector $\check b_1$:

(7.40)¶\[ \check b_1= \Phi^{+} X_1 \]

(To highlight that (7.39) is an approximation, users of DMD sometimes call components of the basis vector $\check b_t = \Phi^+ X_t $ the exact DMD modes.)

Conditional on $X_t$, we can compute our decoded $\check X_{t+j}, j = 1, 2, \ldots $ from either

(7.41)¶\[ \check X_{t+j} = \Phi \Lambda^j \Phi^{+} X_t \]

or use the approximation

(7.42)¶\[ \hat X_{t+j} = \Phi \Lambda^j (W \Lambda)^{-1} \tilde U^T X_t . \]

We can then use $\check X_{t+j}$ or $\hat X_{t+j}$ to forecast $X_{t+j}$.

7.13.3. Using Fewer Modes¶

In applications, we’ll actually want to just a few modes, often three or less.

Some of the preceding formulas assume that we have retained all $p$ modes associated with the positive singular values of $X$.

We can adjust our formulas to describe a situation in which we instead retain only the $r < p$ largest singular values.

In that case, we simply replace $\tilde \Sigma$ with the appropriate $r\times r$ matrix of singular values, $\tilde U$ with the $m \times r$ matrix whose columns correspond to the $r$ largest singular values, and $\tilde V$ with the $n \times r$ matrix whose columns correspond to the $r$ largest singular values.

Counterparts of all of the salient formulas above then apply.

7.14. Source for Some Python Code¶

You can find a Python implementation of DMD here:

https://mathlab.github.io/PyDMD/

7. Singular Value Decomposition (SVD)¶

7.1. Overview¶

7.2. The Setup¶

7.3. Singular Value Decomposition¶

7.4. Properties of Full and Reduced SVD’s¶

7.5. Digression: Polar Decomposition¶

7.6. Principal Components Analysis (PCA)¶

7.7. Relationship of PCA to SVD¶

7.8. PCA with Eigenvalues and Eigenvectors¶

7.9. Connections¶

7.10. Dynamic Mode Decomposition (DMD)¶

7.11. Representation 1¶

7.12. Representation 2¶

7.13. Representation 3¶

7.13.1. Decoder of \(X\) as a linear projection¶

7.13.2. A useful approximation¶

7.13.3. Using Fewer Modes¶

7.14. Source for Some Python Code¶