# Activity 09/16/2020 ## STAT 757 -- Section 1001
Instructor: Colin Grudzien
## Instructions: We will work through the following series of activities as a group and hold small group work and discussions in Zoom Breakout Rooms. Follow the instructions in each sub-section when the instructor assigns you to a breakout room. ## Activities: ### Activity 1: basic properties of matrices #### Question 1: In this problem, we will only assume the following: $$n, p \geq 2$$ and let $\mathbf{X}\in \mathbb{R}^{n \times p}$. Define $\mathbf{X}^\dagger = \left(\mathbf{X}^\mathrm{T} \mathbf{X}\right)^{-1} \mathbf{X}^\mathrm{T}$. Carefully compute the value of the product $\mathbf{X}^\dagger \mathbf{X}.$ ##### Solution to question 1: We remark, while we assume that $\left(\mathbf{X}^\mathrm{T} \mathbf{X}\right)^{-1}$ exists, we do not assume that $\mathbf{X}$ has any well defined inverse. Indeed, if $p\neq n$, then $\mathbf{X}$ will not even be square. Let us define the following matrix $\mathbf{A}\triangleq \left( \mathbf{X}^\mathrm{T} \mathbf{X}\right) \in \mathbb{R}^{p \times p}$. Then by substitution, we find \begin{align} \mathbf{X}^\dagger \mathbf{X} & = \left( \mathbf{X}^\mathrm{T} \mathbf{X}\right)^{-1} \mathbf{X}^\mathrm{T} \mathbf{X} \\ & = \mathbf{A}^{-1} \mathbf{A} \\ & = \mathbf{I} \in \mathbb{R}^{p \times p}. \end{align} by the associativity of matrix multiplication. In particular, we have verified one of the properties of the left pseudo-inverse. These aren't the focus of the course, but they will come up, and it is good exposure to be aware of these objects. They are extremely useful in applied mathematics and statistics for some reasons we will see in upcoming lectures. #### Question 2: We discussed that a projection operator $\Pi$ has the property in general that it is idempotent, i.e., $\Pi^2 = \Pi$. In the special case of the hat matrix, $$\mathbf{H}= \mathbf{X}\left(\mathbf{X}^\mathrm{T} \mathbf{X}\right)^{-1} \mathbf{X}^\mathrm{T},$$ show that $H^2=H$. Then use this fact to show that for the complementary orthogonal projection $$\left(\mathbf{I}-\mathbf{H}\right)^2 = \left(\mathbf{I} - \mathbf{H}\right)$$. ##### Solution: We must take similar care as in the last problem about the existence of inverses. Specifically, we can write \begin{align} \mathbf{H}^2 &= \left[\mathbf{X}\left(\mathbf{X}^\mathrm{T} \mathbf{X}\right)^{-1} \mathbf{X}^\mathrm{T}\right] \left[ \mathbf{X}\left(\mathbf{X}^\mathrm{T} \mathbf{X}\right)^{-1} \mathbf{X}^\mathrm{T}\right] \\ &= \mathbf{X} \left(\mathbf{X}^\mathrm{T} \mathbf{X}\right)^{-1} \left(\mathbf{X}^\mathrm{T} \mathbf{X}\right)\left(\mathbf{X}^\mathrm{T} \mathbf{X}\right)^{-1} \mathbf{X}^\mathrm{T} \end{align} by the associativity of matrix multiplication. Cancelling terms, we recover the definition of the hat matrix directly. Now, we can say \begin{align} \left(\mathbf{I}-\mathbf{H}\right)^2 &= \left(\mathbf{I} - \mathbf{H}\right) \left(\mathbf{I} -\mathbf{H}\right) \\ &=\mathbf{I}^2 - \mathbf{I}\mathbf{H} - \mathbf{H} \mathbf{I} + \mathbf{H}^2 \\ &= \mathbf{I} - \mathbf{H} \end{align} using the property above and the properties of the identity. ### Debrief: We will discuss the result of activity 1 as a class. ### Activity 2: #### Question 1: What is the geometric meaning of the statements \begin{align} \sum_{i=1}^n \hat{\epsilon}_i X_i &= 0 \\ \sum_{i=1}^n \hat{\epsilon}_i\hat{Y}_i &= 0? \end{align} How does this relate to the column span of the design matrix? ##### Solution to question 1: Both of these are statements of orthogonality. Specifically, the residuals are defined as the difference between $\mathbf{Y}$ and the orthogonal projection of $\mathbf{Y}$ into the column span of $\mathbf{X}$, denoted by the fitted values $\hat{\mathbf{Y}}$. The residuals can be written in the form $$\hat{\boldsymbol{\epsilon}} = \left(\mathbf{I} - \mathbf{H}\right)\mathbf{Y}$$ so that they are given precisely by the components of $\mathbf{Y}$ that are orthogonal to the design space. The inner product of $\hat{\boldsymbol{\epsilon}}$ with any vector in the column span of $\mathbf{X}$ must equal to zero by this construction. Because $\hat{\mathbf{Y}}$ lies within the column span of $\mathbf{X}$ by its definition, $$\hat{\mathbf{Y}} = \mathbf{X} \hat{\boldsymbol{\beta}},$$ $\hat{\boldsymbol{\epsilon}}$ must be orthogonal to this vector, as well as the vector of predictor values. #### Question 2: Suppose we are given $n=2$ cases of the data $\left\{\left(X_i,Y_i\right)\right\}_{i=1}^2$. Construct the design matrix for the standard simple regression problem. Do you anticipate issues with performing regression analysis in this data set? Explain why or why not. ##### Solution to question 2: If we are given two cases of data, our design matrix is given as \begin{align} \mathbf{X} = \begin{pmatrix} 1 & X_1 \\ 1 & X_2 \end{pmatrix} \end{align} where the first column corresponds to the intercept term. This situation is troubling because we have $n=2$ cases of data and $p=2$ parameters to estimate. This means that we will not have an estimate for the variance $\sigma^2$ and we are running a risk of overfitting the data. ### Debrief: We will discuss the results of activity 1 as a class. ### Activity 3: #### Question 1: Use the standard model in matrix form and the definition of the least squares estimated parameters, \begin{align} \mathbf{Y} = \mathbf{X} \boldsymbol{\beta} + \boldsymbol{\epsilon} & & \hat{\boldsymbol{\beta}} \triangleq \left(\mathbf{X}^\mathrm{T} \mathbf{X} \right)^{-1} \mathbf{X}^\mathrm{T}\mathbf{Y}, \end{align} to prove that $\hat{\boldsymbol{\beta}}$ is an unbiased estimate of $\boldsymbol{\beta}$. Furthermore, use the definition of the covariance of a random vector, \begin{align} cov(\mathbf{Y})= \mathbb{E}\left\{\left(\mathbf{Y} - \mathbb{E}\left[\mathbf{Y}\right]\right)\left(\mathbf{Y} - \mathbb{E}\left[\mathbf{Y}\right]\right)^\mathrm{T} \right\}, \end{align} to derive the covariance of this estimate.
#### Solution: We note that by definition, \begin{align} \mathbb{E}\left[ \hat{\boldsymbol{\beta}}\right] &= \mathbb{E}\left[\left(\mathbf{X}^\mathrm{T}\mathbf{X}\right)^{-1}\mathbf{X}^\mathrm{T} \mathbf{X} \boldsymbol{\beta} +\left(\mathbf{X}^\mathrm{T}\mathbf{X}\right)^{-1}\mathbf{X}^\mathrm{T}\boldsymbol{\epsilon}\right] \\ & = \mathbb{E}\left[ \mathbf{I} \boldsymbol{\beta} \right] + \mathbb{E}\left[ \left(\mathbf{X}^\mathrm{T}\mathbf{X}\right)^{-1}\mathbf{X}^\mathrm{T}\boldsymbol{\epsilon}\right] \\ &= \boldsymbol{\beta} \end{align} as $\boldsymbol{\epsilon}$ is mean zero. Using the above fact, we can write, \begin{align} cov\left(\hat{\boldsymbol{\beta}}\right) &= \mathbb{E}\left[ \left( \hat{\boldsymbol{\beta}} - \boldsymbol{\beta}\right)\left( \hat{\boldsymbol{\beta}} - \boldsymbol{\beta}\right)^\mathrm{T}\right] \\ &= \mathbb{E}\left[\left\{ \boldsymbol{\beta} - \boldsymbol{\beta} + \left(\mathbf{X}^\mathrm{T}\mathbf{X}\right)^{-1}\mathbf{X}^\mathrm{T}\boldsymbol{\epsilon} \right\} \left\{ \boldsymbol{\beta} - \boldsymbol{\beta} + \left(\mathbf{X}^\mathrm{T}\mathbf{X}\right)^{-1}\mathbf{X}^\mathrm{T}\boldsymbol{\epsilon} \right\}^\mathrm{T}\right] \\ &= \mathbb{E}\left[\left(\mathbf{X}^\mathrm{T}\mathbf{X}\right)^{-1}\mathbf{X}^\mathrm{T}\boldsymbol{\epsilon}\boldsymbol{\epsilon}^\mathrm{T} \mathbf{X} \left(\mathbf{X}^\mathrm{T}\mathbf{X}\right)^{-1}\right]\\ &= \left(\mathbf{X}^\mathrm{T}\mathbf{X}\right)^{-1}\mathbf{X}^\mathrm{T}\left(\sigma^2\mathbf{I}\right)\mathbf{X} \left(\mathbf{X}^\mathrm{T}\mathbf{X}\right)^{-1} \\ &= \sigma^2 \left(\mathbf{X}^\mathrm{T}\mathbf{X}\right)^{-1} \end{align} by canceling terms in the final line, and passing the scalar in front of the equation.