Skip to main content

MATH 345: Linear Algebra and Optimization

Section C.5 Unit 5 applied, geometric, and computational interpretation

These solution sketches correspond to the additional applied and computational interpretation exercises at the end of Unit 5.

Subsection Critical points are candidates

We have
\begin{equation*} \nabla f(x,y)= \begin{bmatrix} 2x\\ -2y \end{bmatrix}. \end{equation*}
Thus the only critical point is \((0,0)\text{.}\)
Along the \(x\)-axis,
\begin{equation*} f(t,0)=t^2, \end{equation*}
which is positive for \(t\ne 0\text{.}\) Along the \(y\)-axis,
\begin{equation*} f(0,t)=-t^2, \end{equation*}
which is negative for \(t\ne 0\text{.}\) Nearby points therefore occur both above and below \(f(0,0)=0\text{.}\) The critical point is a saddle point, not a local minimum or local maximum.

Subsection Directional derivative as a dot product

Use
\begin{equation*} D_{\mathbf{v}}f(\mathbf{a})=\nabla f(\mathbf{a})\cdot \mathbf{v}. \end{equation*}
Then
\begin{equation*} D_{\mathbf{v}_1}f(\mathbf{a})=6, \qquad D_{\mathbf{v}_2}f(\mathbf{a})=8. \end{equation*}
Also,
\begin{equation*} D_{\mathbf{v}_3}f(\mathbf{a}) = \begin{bmatrix} 6\\ 8 \end{bmatrix} \cdot \frac15 \begin{bmatrix} 3\\ 4 \end{bmatrix} = \frac{18+32}{5}=10, \end{equation*}
and
\begin{equation*} D_{\mathbf{v}_4}f(\mathbf{a})=-10. \end{equation*}
The largest listed directional derivative occurs in direction \(\mathbf{v}_3\text{,}\) and the smallest occurs in direction \(\mathbf{v}_4\text{.}\) This is the Unit 1 dot product measuring alignment: the derivative is largest when the direction is aligned with the gradient and smallest when it points opposite the gradient.

Subsection Reading a gradient descent loop

The loop represents
\begin{equation*} \mathbf{x}_{k+1}=\mathbf{x}_k-\alpha\nabla f(\mathbf{x}_k). \end{equation*}
The learning rate is alpha. The vector grad_f(x) must have the same shape as x. Replacing the minus sign by a plus sign would move in the direction of steepest increase instead of steepest decrease. The code does not prove that a minimum has been found. It only describes the update rule; convergence depends on the function, starting point, learning rate, and stopping rule.

Subsection Learning-rate diagnosis

The learning rate \(\alpha=0.05\) is making slow but steady progress. The learning rate \(\alpha=0.20\) is making faster useful progress. The learning rate \(\alpha=1.05\) appears unstable because the loss increases.
A decreasing loss table does not prove that the global minimum has been found. It only reports what happened for the displayed iterates.

Subsection Eigenvectors as special directions

We compute
\begin{equation*} A\mathbf{v}_1= \begin{bmatrix} 3\amp1\\ 0\amp2 \end{bmatrix} \begin{bmatrix} 1\\ 0 \end{bmatrix} = \begin{bmatrix} 3\\ 0 \end{bmatrix} = 3\mathbf{v}_1. \end{equation*}
So \(\mathbf{v}_1\) is an eigenvector with eigenvalue \(3\text{.}\)
Next,
\begin{equation*} A\mathbf{v}_2= \begin{bmatrix} 3\amp1\\ 0\amp2 \end{bmatrix} \begin{bmatrix} -1\\ 1 \end{bmatrix} = \begin{bmatrix} -2\\ 2 \end{bmatrix} = 2 \begin{bmatrix} -1\\ 1 \end{bmatrix} = 2\mathbf{v}_2. \end{equation*}
So \(\mathbf{v}_2\) is an eigenvector with eigenvalue \(2\text{.}\) The matrix map stretches the \(\mathbf{v}_1\)-direction by a factor of \(3\) and the \(\mathbf{v}_2\)-direction by a factor of \(2\text{.}\)

Subsection Quadratic form and definiteness

For the three listed vectors,
\begin{equation*} \begin{bmatrix} 1\\ 0 \end{bmatrix}^T H \begin{bmatrix} 1\\ 0 \end{bmatrix} =4, \end{equation*}
\begin{equation*} \begin{bmatrix} 0\\ 1 \end{bmatrix}^T H \begin{bmatrix} 0\\ 1 \end{bmatrix} =1, \end{equation*}
and
\begin{equation*} \begin{bmatrix} 1\\ 1 \end{bmatrix}^T H \begin{bmatrix} 1\\ 1 \end{bmatrix} =5. \end{equation*}
For a general vector
\begin{equation*} \mathbf{h}= \begin{bmatrix} h_1\\ h_2 \end{bmatrix}, \end{equation*}
\begin{equation*} \mathbf{h}^TH\mathbf{h}=4h_1^2+h_2^2. \end{equation*}
This is positive for every nonzero \(\mathbf{h}\text{,}\) so \(H\) is positive definite. If \(H\) is the Hessian at a critical point, the second derivative test gives a local minimum.

Subsection Hessian eigenvalue classification

If the Hessian eigenvalues are \(3,5\text{,}\) the Hessian is positive definite, so the critical point is a local minimum.
If the Hessian eigenvalues are \(-2,-7\text{,}\) the Hessian is negative definite, so the critical point is a local maximum.
If the Hessian eigenvalues are \(4,-1\text{,}\) the Hessian is indefinite, so the critical point is a saddle point.
If the Hessian eigenvalues are \(0,2\text{,}\) the test is inconclusive. The zero eigenvalue means the quadratic approximation has a flat direction, so the second-order term alone does not decide the local behavior.

Subsection Least squares from gradients

We compute
\begin{equation*} A^TA= \begin{bmatrix} 3\amp3\\ 3\amp5 \end{bmatrix}, \qquad A^T\mathbf{b}= \begin{bmatrix} 5\\ 6 \end{bmatrix}. \end{equation*}
The normal equations are
\begin{equation*} A^TA\hat{\mathbf{x}}=A^T\mathbf{b}, \end{equation*}
or
\begin{equation*} \begin{bmatrix} 3\amp3\\ 3\amp5 \end{bmatrix} \hat{\mathbf{x}} = \begin{bmatrix} 5\\ 6 \end{bmatrix}. \end{equation*}
Solving gives
\begin{equation*} \hat{\mathbf{x}}= \begin{bmatrix} 7/6\\ 1/2 \end{bmatrix}. \end{equation*}
Then
\begin{equation*} A\hat{\mathbf{x}}= \begin{bmatrix} 7/6\\ 5/3\\ 13/6 \end{bmatrix}, \end{equation*}
so the residual is
\begin{equation*} \mathbf{r} = \mathbf{b}-A\hat{\mathbf{x}} = \begin{bmatrix} -1/6\\ 1/3\\ -1/6 \end{bmatrix}. \end{equation*}
Finally,
\begin{equation*} A^T\mathbf{r} = \begin{bmatrix} 0\\ 0 \end{bmatrix}. \end{equation*}
This is the residual-orthogonality condition from Unit 4 and the gradient-zero condition from Unit 5.

Subsection Residual orthogonality in code

The vector xhat is a least-squares coefficient vector. The vector r is the residual
\begin{equation*} \mathbf{r}=\mathbf{b}-A\hat{\mathbf{x}}. \end{equation*}
The expression A.T @ r checks whether the residual is orthogonal to the columns of \(A\text{.}\) It can be close to zero even when r is not close to zero. Unit 4 used this residual-orthogonality condition to derive the normal equations. Unit 5 obtains the same equations by setting the gradient of \(\|A\mathbf{x}-\mathbf{b}\|^2\) equal to zero.

Subsection Fixed-hidden-layer design matrix

For \(t=-1,0,1\text{,}\) the design matrix is
\begin{equation*} A= \begin{bmatrix} 1\amp\tanh(-1)\amp\tanh(-2)\amp\tanh(0)\\ 1\amp\tanh(0)\amp\tanh(-1)\amp\tanh(1)\\ 1\amp\tanh(1)\amp\tanh(0)\amp\tanh(2) \end{bmatrix}. \end{equation*}
The trained vector is
\begin{equation*} \mathbf{c}= \begin{bmatrix} c_0\\ c_1\\ c_2\\ c_3 \end{bmatrix}. \end{equation*}
This is a least-squares problem because the prediction vector is \(A\mathbf{c}\text{,}\) and training chooses \(\mathbf{c}\) so that \(A\mathbf{c}\) is close to \(\mathbf{y}\text{.}\) The model is not linear as a function of \(t\text{,}\) because it contains \(\tanh(t)\text{,}\) \(\tanh(t-1)\text{,}\) and \(\tanh(t+1)\text{.}\) It is linear as a function of \(\mathbf{c}\text{,}\) which is why least squares applies.

Subsection Rank-one update

The outer product is
\begin{equation*} \mathbf{g}\mathbf{h}^T = \begin{bmatrix} 1\\ -2\\ 3 \end{bmatrix} \begin{bmatrix} 4\amp0 \end{bmatrix} = \begin{bmatrix} 4\amp0\\ -8\amp0\\ 12\amp0 \end{bmatrix}. \end{equation*}
Its shape is \(3\times 2\text{.}\) Since every column is a scalar multiple of \(\mathbf{g}\text{,}\) the column space has dimension at most one, so the rank is at most one.
The update
\begin{equation*} W_{\mathrm{new}}=W-\alpha \mathbf{g}\mathbf{h}^T \end{equation*}
changes \(W\) by subtracting a scaled rank-one matrix. In NumPy, np.outer(g, h) is appropriate when g and h are stored as one-dimensional arrays because it explicitly forms the matrix \(\mathbf{g}\mathbf{h}^T\text{.}\)