# Proof of normal equation in regression using tensor notation

I'm struggling with a proof of the normal equation, so I posted a question which hopefully will get resolved soon. However, I mentioned there that I'm uncomfortable with the proofs dealing with matrix calculus, particularly when it comes to derivatives of product of vectors or product of matrices and vectors. Furthermore, there are a couple of "layout conventions" which are not even used consistently. Because of this, I thought I could get rid of all that insanity by using tensor notation instead.

In relation to the normal equation question, I would be very interested in looking at a proof of this equation (here you can see the context):

$$\beta = (X^{T}X)^{-1}X^{T}{\bf{y}}$$

using tensor notation.

And as a more general request, I'd like to get an advice on the issue of dealing with matrix calculus and derivatives. Particularly, I want to hear about your approach when you have to calculate something of this kind given the inconsistencies in notations and non-intuitive nature of some identities. Do you consult Wikipedia or some handbook every time you want to calculate a derivative of some product of vectors or matrices? Do you memorize some recurrent identities? Do you know an alternative procedure?

• Offhand I'm familiar with at least a half dozen conventions that could validly be called "tensor notation." One of them is infamously called the "debauch of indices": I hope that's not the one you mean! But what exactly do you want? – whuber Apr 2 '13 at 4:01
• I didn't know there are so many conventions. I had in mind something like this: docs.google.com/… in page 4. – Robert Smith Apr 2 '13 at 4:04
• That document uses at least two notations. Notice how quickly the author abandons the multiple-subscript notation for the simpler and clearer vector-matrix notation. Following that procedure yourself is already pretty good advice--if you get tied up tracking subscripts, you will lose sight of what's really going on. – whuber Apr 2 '13 at 4:22
• That is actually true. I don't know what you mean by "the author abandons the multiple-subscript notation for the simpler and clearer vector-matrix notation", though. Which part are you referring to? – Robert Smith Apr 2 '13 at 4:38
• At the beginning, equation (1)--a partially written-out array--, equation (2)--the "$[a_{ij}]$" notation--, and the symbol "$A$" at the end of the subsequent paragraph all refer to the same mathematical object. That's three notations right there. Throughout the rest of the document, the first notation is used rarely (eqs (24) and (67) only) and occasionally in abbreviated (block-matrix) form, as in equation (19); the second notation is almost as rarely used (as in equation (13)); and the third notation--which has no subscripts--is used hundreds of times. – whuber Apr 2 '13 at 12:34

A "modern" (post-1900) view of tensors is a geometric one. Perhaps the clearest notation would therefore be a figure, because the "normal equations" are nothing more than a familiar, ages-old theorem of plane geometry!

As a point of departure, let's begin with an expression from your previous question. The setting is Ordinary least squares (OLS). The data are the "design matrix" $(x_{i,j})$, $1 \le i \le n$, $1 \le j \le p$ of "independent variables" and a vector of values of the "dependent variable" $y = (y_i)$, $1 \le i \le n$; each dependent value $y_i$ is associated with the $p$-tuple of independent values $(x_{i,1}, x_{i,2}, \ldots, x_{i,p}).$

OLS aims to find a vector $\beta = (\beta_1, \beta_2, \ldots, \beta_p)$ minimizing the sum of squares of residuals

$$\sum_{i=1}^n\left(y_i - \sum_{j=1}^p \left(x_{i,j}\beta_j\right)\right)^2.$$

Geometrically, the inner sum of products represents the application of the linear transformation determined by the matrix $x = (x_{i,j})$ to the $p$-vector $\beta$. As such we conceive of $x$ as representing a map $x: \mathbb{R}^p\to \mathbb{R}^n$. The outer sum is the usual Pythagorean formula for the Euclidean distance. Geometrically, then,

OLS seeks a vector $\hat\beta \in \mathbb{R}^p$ for which $x\hat\beta\in \mathbb{R}^n$ is as close to $y$ as possible.

The test of this is to take any other vector $\beta'\in \mathbb{R}^p$ and to compare distances within the triangle formed by the three points $y$, $x\hat\beta$, and $x\beta'$:

According to Euclid, three (distinct) points determine a plane, so the figure is utterly faithful: all our work is done within this plane.

We also know from Euclid that the ray $y - x\hat\beta$ must be perpendicular to the ray $x\beta' - x\hat\beta$. (This is the heart of the matter; in making this observation, we are really done and all that remains is to write down an equation expressing this perpendicularity.) Modern geometry has several notations for perpendicularity, which involves the (usual) "dot product" or positive-definite linear form (a "tensor of type $(2,0)$" if you wish). I will write it as $\langle\ ,\ \rangle$, whence we have deduced

$$\langle x\beta', y-x\hat\beta\rangle = 0$$

for all $\beta\in\mathbb{R}^p$. Because the usual dot product of two vectors in $\mathbb{R}^n$ is

$$\langle u, v\rangle = u^\intercal v,$$

the preceding may be written

$$\beta'^\intercal x^\intercal (y - x\hat\beta) = 0.$$

Geometrically, this says that the linear transformation from $\mathbb{R}^{p*}\to\mathbb{R}^n$ defined by applying $x^\intercal (y - x\hat\beta)$ to $\beta'^\intercal$ is the zero transformation, whence its matrix representation must be identically zero:

$$x^\intercal (y - x\hat\beta) = 0.$$

We are done; the conventional notation

$$\hat\beta = \left(x^\intercal x\right)^{-}\left(x^\intercal y\right)$$

(where "$\left(\right)^{-}$" represents a generalized inverse) is merely an algebraically suggestive way of stating the same thing. Most computations do not actually compute the (generalized) inverse but instead solve the preceding equation directly.

If, as an educational exercise, you really, really want to write things out in coordinates, then I would recommend adopting coordinates appropriate to this configuration. Choosing an orthonormal basis for $\mathbb{R}^n$ adapted to the plane determined by $y$ and $x\beta'$ for some arbitrary $\beta'$ allows you to forget the remaining $n-2$ coordinates, because they will be constantly $0$. Whence $y$ can be written $(\eta_1, \eta_2)$, $x$ can be written $(\xi_1, \xi_2)$, and $\beta'$ can be considered a number. We seek to minimize

$$\sum_{i=1}^n\left(y_i - \sum_{j=1}^p \left(x_{i,j}\beta_j\right)\right)^2=\left(\eta_1 - \xi_1 \beta\right)^2 + \left(\eta_2 - \xi_2 \beta\right)^2.$$

OLS claims that this is achieved for

$$\hat\beta = \left(x^\intercal x\right)^{-}\left(x^\intercal y\right)=\left(\xi_1^2 + \xi_2^2\right)^{-1}\left(\xi_1 \eta_1 + \xi_2 \eta_2\right)$$

provided $\xi_1^2 + \xi_2^2 \ne 0$. An elementary way to show this is to expand the former expression into powers of $\beta$ and apply the Quadratic Formula.

• Thanks! I was aware of the geometric proof but I'm more interested in a proof actively using tensors in order to get the terms of $\beta$ – Robert Smith Apr 2 '13 at 16:18
• This proof does use tensors! – whuber Apr 2 '13 at 17:28
• Yes, but the tensors are not central to the proof. Basically, I want to look at the proof because I'd like to get a taste of how to deal with the derivative to minimize the error function in tensor notation. – Robert Smith Apr 2 '13 at 17:41
• That might be a question for the math site rather than here. – whuber Apr 2 '13 at 17:43
• It could be but I have found that here I get more relevant answers on these topics than in math.stackexchange – Robert Smith Apr 2 '13 at 17:53