When consider the following DGP : $y=X\beta^{*}+\epsilon$ where $\beta^{*}$ is a $\tilde k\times1\ $ vector.

Define the projection matrices: $P_{X}=X(X^{T}X)^{-1}X^{T}$ and $M_{X}=I-X(X^{T}X)^{-1}$.

first problem is prove that $X^{T}P_{X}=X^{T}$ and $ X^{T}M_{X}=0$

I thought this problem is too simple to think because P is just I. So, I think that there is an intention for given problem considering second problem.

Second problem is Prove thath if $X_{(i)}$ is the $i^{th}$ column of $X$, then, $M_{X}X_{(i)}=0$

My question is

  1. What is the point of first problem?

  2. Can I solve these problem by considering $X_{(i)}=XA\ $ where A has a vector consisting of just 1 and the other columns are all zero? or Should I utilize first problem's solution?

  • $\begingroup$ Please post separate questions separately. $\endgroup$
    – Giskard
    Commented Oct 14, 2017 at 10:43
  • 1
    $\begingroup$ The "first problem" is indeed immediately proven by simply inserting the expression for the projection matrices. For the second problem, use the fact that $M_X$ is symmetric. $\endgroup$ Commented Oct 14, 2017 at 12:04

1 Answer 1


This kind of projections in econometrics are usually employed for partialling out some covariates from a linear regression.

Observe that in general $P_X\neq I$. Consider $X=\begin{bmatrix}1&5\\1&0\\1&1\end{bmatrix}$. Then $X'X=\begin{bmatrix}3&6\\6&26\end{bmatrix}$ and $(X'X)^{-1}=\begin{bmatrix}26/42&-6/42\\-6/42&3/42\end{bmatrix}$ then $X(X'X)^{-1}X^T=\frac{1}{42}\begin{bmatrix}41&-4&5\\-4&26&20\\5&20&17\end{bmatrix}\neq I$.

First problem: $X^TP_X=X^TX(X'X)^{-1}X^T=(X^TX)(X'X)^{-1}X^T=I_{(\tilde k\times\tilde k)}X^T=X^T$

Similarly: $X^TM_X=X^T[I_{(\tilde k\times\tilde k)}-P_X]=X^TI_{(\tilde k\times\tilde k)}-X^TP_X=X^T-X^T=0$

The second problem is a direct consequence of $X^TM_X=0$, simply now you can post-multiply for $X_{(i)}$ since you are considering just one single column. The approach of $X_{(i)}=XA$ where $A$ is a vector $\tilde k\times 1$ of 0 apart from the $i$-th element which is 1. Then it should be straightforward to see: $0=(XA)^TM_X=M_XX_{(i)}$

  • $\begingroup$ I have never thought that kind of solving. It was very confusing for me to utilize such a matrix. Thanks a lot! $\endgroup$
    – Jeffrey
    Commented Oct 14, 2017 at 14:11
  • $\begingroup$ @GabMac Are there any econometrics books that present this linear regression thinking? $\endgroup$
    – Linkman
    Commented Mar 14, 2019 at 23:10

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.