Meetup 2: Least Squares

George Hagstrom

Week Summary

  • Starting on Least Squares
  • Chapter 12 of VMLS
  • Keep working on Lab 1, due this Sunday at midnight

System of Linear Equations

  • When can we solve: \[ A\mathbf{x} = \mathbf{b}\,?\]
  • \(A\) is an \(m\times n\) matrix
  • \(\mathbf{x}\) is an \(n\)-vector
  • \(\mathbf{b}\) is an \(m\)-vector

Tall, Wide, and Square Matrices

  • \(m=n\) square

  • Exact solution if rows/columns are independent \[ A = \begin{pmatrix} a_{11} & \cdots & \cdots & a_{1n} \\ \vdots & \cdots & \cdots & \vdots \\ \vdots & \cdots & \cdots & \vdots \\ a_{n1} & \cdots & \cdots & a_{nn} \end{pmatrix} \]

  • \(\mathbf{x} = A^{-1}\mathbf{b}\)

Tall, Wide, and Square Matrices

  • \(m<n\) wide
  • Generally multiple solutions because dependent columns \[ A = \begin{pmatrix} a_{11} & \cdots & \cdots & \cdots & a_{1n} \\ \vdots & \cdots & \cdots & \cdots & \vdots \\ a_{m1} & \cdots & \cdots & \cdots & a_{mn} \end{pmatrix} \]
  • Common ML situation (more parameters than data points)
  • Particular solution plus the null space

Tall, Wide, and Square Matrices

  • \(m>n\) tall
  • Generally no exact solution \[ A = \begin{pmatrix} a_{11} & \cdots & a_{1n} \\ \vdots & \cdots & \vdots \\ \vdots & \cdots & \vdots \\ \vdots & \cdots & \vdots \\ a_{m1} & \cdots & a_{mn} \end{pmatrix} \]

Inspiration for Least Squares

  • Can’t solve? Minimize the error \[ \mathbf{r} = A\mathbf{x}-\mathbf{b}\]
  • Optimization problem is: \[\textrm{Find:}\quad \min_{\mathbf{x}} \|A\mathbf{x} - \mathbf{b}\|^2\]
  • The \(\mathbf{x}\) that minimizes this objective (the squared residual) is the least squares solution
  • Linear regression

How to Solve It?

  • Take derivative and set to 0: \[ \frac{\partial}{\partial \mathbf{x}}\|A\mathbf{x}-\mathbf{b}\|^2 = 2A^T A\left(\mathbf{x} - A^T\mathbf{b}\right) \]
  • Two ways to get that: www.matrixcalculus.org

How to Solve It?

  • Take derivative and set to 0: \[ \frac{\partial}{\partial \mathbf{x}}\|A\mathbf{x}-\mathbf{b}\|^2 = 2A^T A\left(\mathbf{x} - A^T\mathbf{b}\right) \]
  • Two ways to get that

How to Solve It?

  • Get System of Linear Equations: \[ A^TA\mathbf{x} = A^T\mathbf{b} \]
  • \(A^TA\) is an \(n\times n\) symmetric matrix called the Gram matrix
  • Invertible if columns of \(A\) are independent:

\[ \mathbf{x} = \left(A^T A\right)^{-1}A^T\mathbf{b}\]

  • This is the exact formula for the least squares solution

Properties

  • Morse-Penrose Pseudoinverse: \[A^+ = (A^TA)^{-1}A^T\]
  • \(n\times m\) matrix
  • If \(A\) is invertible, \(A^+ = A^{-1}\)

Applications: Linear Regression

  • Suppose you have \(m\) observations, each with \(n\) features

  • \(\mathbf{a}_{i}\) is \(i\)th obs

  • Want to predict the variable \(\mathbf{b}\) using an Affine function: \[ \mathbf{a}_i^T\mathbf{x} +c = b_i \]

  • or

    \[ \sum_{j=1}^m a_{ij} x_j + c = b_i \]

Linear Regression

  • Augmented matrix \(A\) and coefficients \(\mathbf{x}\): \[ A = \begin{pmatrix} a_{11} & \cdots & a_{1n} & 1 \\ \vdots & \cdots & \vdots & 1 \\ a_{m1} & \cdots & a_{mn} & 1 \end{pmatrix} \] \[ \mathbf{x} = \begin{pmatrix} x_1 & x_2 & \cdots & x_n & c \end{pmatrix}^T \]
  • Now get standard least squares problem:

\[ \min_{\mathbf{x}}\|A\mathbf{x}-\mathbf{b}\|^2\]

Applications: Marketing Budget

  • You are tasked with spending a marketing budget
  • There are \(10\) different demographic groups to target
Group
1 18-25, play sports
2 26-30, play sports
3 31-35, play sports
4 36-40, play sports
5 41-45, play sports
6 18-25, no sports
7 26-30, no sports
8 31-35, no sports
9 36-40, no sports
10 41-45, no sports

Applications: Marketing Budget

  • You are tasked with spending a marketing budget
  • There are \(10\) different demographic groups to target
  • 4 different marketing channels
Channel
1 Instagram
2 Wine Mag
3 Facebook
4 UFC

Applications: Marketing Budget

  • Views: $1 on channel \(i\) gets \(R_{ij}\) views in group \(j\)
   Instagram  Wine Mag  Facebook   UFC      Demo
0        3.0       0.1       0.4  6.00   18-25 S
1        2.5       0.3       0.6  6.00   26-30 S
2        2.3       0.4       1.0  5.50   31-35 S
3        1.8       0.6       1.3  5.00   36-40 S
4        1.0       0.6       2.0  4.50   41-45 S
5        3.3       0.8       0.5  0.20  18-25 NS
6        2.6       1.1       0.8  0.12  26-30 NS
7        2.2       1.3       1.1  0.10  31-35 NS
8        1.4       2.3       1.3  0.40  36-40 NS
9        1.2       4.2       1.7  0.20  41-45 NS

Applications: Marketing Budget

  • Views: $1 on channel \(i\) gets \(R_{ij}\) views in group \(j\)

Marketing Budget

  • Problem: spend money to get \(\mathbf{v}_{dem}\) impressions
  • Decision variable \(\mathbf{s}\) is spending in each channel
  • Solve least squares problem: \[ \min_{\mathbf{s}} \|R\mathbf{s} - \mathbf{v}_{dem}\|^2 \]
  • \(R\mathbf{s}\) is views

Marketing Budget

  • Suppose \(\mathbf{v}_{dem}\) is uniform
v_dem = np.ones(10)

s = np.linalg.lstsq(R,v_dem)[0]

s_dat = pd.DataFrame(s, columns = ["Spending"])
s_dat["Channel"] = ['Instagram','Wine Mag','Facebook','UFC'] 

(
  ggplot(s_dat,aes(x="Channel",y="Spending"))
  + geom_point(size=5) 
  + theme_bw(base_size=14)
)

Marketing Budget

  • Suppose \(\mathbf{v}_{dem}\) is uniform

Marketing Budget

  • Suppose \(\mathbf{v}_{dem}\) is uniform

Marketing Budget

  • Suppose we want to target only 41-45 Non-Sports Players

Marketing Budget

  • Suppose we want to target only 41-45 Non-Sports Players

Marketing Budget Issues

  • Need Constraints
    • No negative money in adds: \(\mathbf{s} \geq 0\)
    • No infinite budget: \(\sum_{i}s_i \leq s_{max}\)
  • Need a better objective function
    • Weird to target precise \(\mathbf{v}_{dem}\)
    • Better to have different value for each ad impression

Network Tomography

  • Urban Transit
  • Internet
  • Know when people enter and leave
  • Know path they took
  • How long does it take to cross each link?

Network Tomography

Network Tomography

Network Tomography

Network Tomography

  • For trip \(j\) vector \(\mathbf{r}_j\)
    • \(r_{ij} = 1\) if \(i\)th “link” traversed on \(j\)th trip
    • \(r_{ij} = 0\) otherwise
    • Matrix \(R\) contains \(r_{ij}\)
  • For \(t_j\) is the time taken on trip \(j\)
  • Decision variables \(d_i\) is the time taken on the \(i\)th link
  • Time is \(\sum r_{ij}d_i\)

Network Tomography

  • Minimize Least Squares Error:

\[ \textrm{Find:}\quad \min_{\mathbf{d}} \|R\mathbf{d} - \mathbf{t}\|^2 \]

  • Many similar problems in optimization theory
  • This one is the simplest

Numerical Issues With Least Squares

  • Suppose you need to solve a system of linear equations: \[ A\mathbf{x} = \mathbf{b} \]

  • Which of these is better?

    • np.linalg.inv(A) @ b
    • np.linalg.solve(A, b)

Matrix Inverses

  • Consider Singular Value Decomposition of \(A\): \[ A = U\Sigma V^T \]
  • \(U\) and \(V\) are orthonormal matrices
  • \(\Sigma\) contains the singular values \[ \Sigma = \begin{bmatrix} \sigma_{1} & & \\ & \ddots & \\ & & \sigma_{n} \end{bmatrix} \]

Condition Number and Accuracy

  • Accuracy of np.linalg.inv(A) @ b related to \(\frac{\sigma_{max}}{\sigma_{min}}\)
  • This is called the “condition number” of the matrix \(A\)
%timeit np.linalg.inv(A) @ b
%timeit np.linalg.solve(A,b)

x1 = np.linalg.inv(A) @ b
x2 = np.linalg.solve(A,b)

print("Matrix Inverse Error: ",np.linalg.norm(A @ x1 - b))
print("Linear Solve Error: ",np.linalg.norm(A @ x2 - b))
6.04 ms ± 937 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
1.56 ms ± 374 µs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)
Matrix Inverse Error:  4.48934571429838e-07
Linear Solve Error:  2.344689110789182e-15
  • 100 million times more accurate, 5 times faster

Issue is worse for \(A^TA\)

  • What about \((A^TA)^{-1}\)?
  • Let’s look at SVD: \[ A^TA = V \Sigma^T\Sigma U^T\] \[ \Sigma^T\Sigma = \begin{bmatrix} \sigma_{1}^2 & & \\ & \ddots & \\ & & \sigma_{n}^2 \end{bmatrix} \]
  • Condition number problems 💀 💀 💀 💀

Do Ill-Conditioned Matrices Occur?

  • Very Common
  • Collinear predictors in linear regression
  • Main issue in Deep Learning
  • Noisy Inverse Problems
  • Some of the solutions are basically just regularization

Alternatives

  • Software packages generally have routines to solve least squares without computing ill-conditioned matrix inverse
  • Good example is \(A = QR\) factorization
  • Here \(Q\) is orthogonal and \(R\) is upper triangular
  • Then \[ A^+ = (A^TA)^{-1}A^T = R^{-1} Q \]
  • In practice np.linalg.linsolve(R,Q @ b)

What if \(A\) is really big?

  • Constructing \(A^TA\) costs \(mn^2\) operations
  • Solving \(A^TA\mathbf{x} = A^T\mathbf{b}\) costs \(n^3\) operations
  • \(A^TA\) could be a TB or larger…..
  • Solution 1- variety of iterative methods (HW)
  • Solution 2- Randomized Algorithms (Random Kaczmarz)

Thanks!

DATA 609