If we have 100 data points and want to perform polynomial regression, the maximum degree of our polynomial is n-1, where n is the number of data points. In this case, the maximum degree would be 99. I am aware that this will almost always overfit the data, but my question is this: why can we only go up to 99? What is to stop us from taking a polynomial of degree, say, 14 million, and fitting the data?

I am aware of the n-1 bound, but haven't found a good explanation online as to why this bound exists.

  • 2
    $\begingroup$ A simple, intuitive, smallest-case explanation here $\endgroup$
    – Glen_b
    Oct 12 '14 at 1:15

In short, what you are doing to fit a polynomial regression model like \begin{align*} Y&=\beta_0+\beta_1X+\beta_2X^2+\cdots+\beta_pX^p+\varepsilon\\ &=\mathbf{X}\boldsymbol\beta+\varepsilon \end{align*} is to perform a linear regression model with $p$ variables given by $(X,X^2,\ldots,X^p)$, so your design matrix, from a sample $\{(X_i,Y_i)\}_{i=1}^n$ is $$ \mathbf{X}=\pmatrix{\begin{array}{ccc}1 & X_1 & \cdots & X_1^p\\ \vdots & \vdots &\ddots &\vdots\\ 1 & X_n & \cdots & X_n^p \end{array}}_{n\times(p+1)}. $$ The estimate for $\boldsymbol\beta$ is $$ \hat{\boldsymbol\beta}=(\mathbf{X}^T\mathbf{X})^{-1}\mathbf{X}\mathbf{y}, $$ with $\mathbf{y}=(Y_1,\ldots,Y_n)^T$.

The important point is that $\mathbf{X}^T\mathbf{X}$ is a $(p+1)\times(p+1)$ matrix and that $\mathrm{rank}(\mathbf{X}^T\mathbf{X})=\mathrm{rank}(\mathbf{X})$. Then, if $p=n-1$, $\mathbf{X}^T\mathbf{X}$ has dimension $n\times n$ and its rank is $n$, so no problem, is invertible. But if $p=n$, the dimension of $\mathbf{X}^T\mathbf{X}$ is $(n+1)\times(n+1)$ and the rank remains $n$, so in that case (and also if $p>n$) is not invertible. In other words, linear dependence arises as the result of fitting more variables than sample points you have.


Suppose you are fitting a 1-degree polynomial to 2 points- i.e., a line through 2 points. There is one and only one answer. If you fit (least squares) a line through 3,...,n points, there is also only one line, although the line can't go through all the points, it just finds a 'best place' for fitting the line among the points (again, least squares).

Now suppose you want to fit a quadratic (2-degree polynomial) to those same 2 points. Many quadratics will fit those two points exactly- an infinite number, in fact. Many cubics will also, many quartics, etc. So there isn't one solution, there are infinitely many solutions in all those cases. That's why your equation-solving software spits up at that point. n points can be fitted exactly by an n-1 degree polynomial, and in an approximate (least squares again) way by a polynomial of degree 1,...,n-2. But a polynomial of degree n will fit all those n points exactly, but with 'wiggle room left over'- there will be infinitely many degree n polynomials which will fit. To decide on which one of those infinitely many solutions you would want, you'd need to have some additional criteria besides just the points. The higher the polynomial degree, the more criteria you would need. And your typical fitting software would have to work differently. This begins to get a bit complex.

(But for the most part, you don't want a really high-degree polynomial to fit your data anyway- you're hoping you can explain a lot by fairly simple equations)


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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