Least Squares

Definition and Derivations

We have already spent much time finding solutions to

Ax  =  b

If there isn't a solution, we attempt to seek the x that gets closest to being a solution.  The closest such vector will be the x such that

Ax  =  projWb

where W is the column space of A. Notice that b - projWb is in the orthogonal complement of W hence in the null space of AT.  Hence if x is a this closest vector, then

AT(b - Ax)  =  0        ATAx  =  ATb

Now we need to show that ATA nonsingular so that we can solve for x.

Lemma

If A is an m x n matrix of rank n, then ATA is nonsingular.

Proof

We want to show that the null space of ATA is zero.  If

0  =  ATAx

then multiplying both sides by xT, we get

0  =  xTATAx  =  (Ax)TAx  =  Ax . Ax  =  ||Ax||2

If the magnitude of a vector is zero, then the vector is zero, hence

Ax  =  0

Since

rank(A)  =  n

we can conclude that

x  =  0

We can now state the main theorem.

Theorem

Let A be an m x n matrix or rank n, then the system

Ax  =  b

has the unique least squares solution

x  =  (ATA) -1ATb

Examples

Example

Find the least squares solution to

Ax  =  b

with Solution

We can quickly check that A has rank 2 (the first two rows are not multiples of each other).  Hence we can compute Notice that not exactly b, but as close as we are going to get.

Least Squares Regression Line

Of fundamental importance in statistical analysis is finding the least squares regression line.

Example

An engineer is tracking the friction index over mileage of a breaking system of a vehicle.  She expects that the mileage-friction relationship is approximately linear.  She collects five data points that are show in the table below.

 Mileage 2000 6000 20,000 30,000 40,000 Friction Index 20 18 10 6 2

The graph below shows these points We are interested in the line that best fits the data.  More specifically, if b is the vector of friction index data values and y is the vector consisting of y values when we plug in the mileage data for x and find y by the equation of the line, then we want the line that minimizes the distance between b and y.  If the equation of the line is

ax + b  =  y

then we get the five equations

2a + b  =  20
6a + b  =  18
20a + b  =  10
30a + b  =  6
40a + b  =  2

The corresponding matrix equation is

Ax  =  b

or Although this does not have an exact solution, it does have a closest solution.  We have We can conclude that the equation of the regression line is

y  =  -0.48x + 20.6

Best Fitting Curves

Often, a line is not the best model for the data.  Fortunately the same technique works if we want to use other nonlinear curves to fit the data.  Here we will explain how to find the least squares cubic.  The process for other polynomials is similar.

Example

A bioengineer is studying the growth of a genetically engineered bacteria culture and suspects that is it approximately follows a cubic model.  He collects six data points listed below

 Time in Days 1 2 3 4 5 6 Grams 2.1 3.5 4.2 3.1 4.4 6.8

He assumes the equation has the form

ax3 + bx2 + cx + d  =  y

This gives six equations with four unknowns

a +     b +   c + d  =  2.1
8a +   4b + 2c + d  =  3.5
27a +   9b + 3c + d  =  4.2
64a + 16b + 4c + d  =  3.1
125a + 25b + 5c + d  =  4.4

216a + 36b + 6c + d  =  6.8

The corresponding matrix equation is We can use the least squares equation to find the best solution So that the best fitting cubic is

y  =  0.2x3 - 2.0x2 + 6.1x - 2.3

The graph is shown below 