Least squares approximation question

I am taking a course on scientific computation, and we just went over least squares approximation. My question is specifically about approximating using polynomials. I understand that if you have n+1 data points, you can find a unique polynomial of degree n that describes all of these points. But I can also see why this is not always ideal. You can get a lot of noise in between data points with such an approach. I suppose it's nice to get a lower degree polynomial that estimates your data well enough.

My question is: how do you decide in practice what degree of polynomial you are going to use? Is there a rule of thumb, or does it depend solely on the problem at hand? Do we have to take into account various tradeoffs when deciding between more or less degrees? Or am I misunderstanding something here?

• I think in practice people use things like spline interpolation en.wikipedia.org/wiki/Spline_interpolation so that low order polys are used, but they fit well with each others over the overall domain. This way one does not have to guess for an overall polynomial order. – Nasser Mar 23 '14 at 3:14
• Thanks for the link. We haven't gone over splines yet, so this is interesting reading. – Uday Pramod Mar 23 '14 at 3:33
• What is it exactly that you want to do? Are you trying to interpolate the points or fit given data? For example, it is useless to interpolate data that consists of a normal distribution with noise. For the former, Nasser's answer is good. For the latter, the fit function depends solely on the problem at hand and is in many cases not polynomial. – hauntergeist Mar 23 '14 at 15:58
• You might be interested in the answers of this question on cross validated. – Bort Jan 10 at 21:39

The most important aspect of interpolation and curve fitting is to understand why high order polynomial fits can be an issue and what the other options are and then you can understand when they are/are not a good choice.

A few issues with high order polynomials:

• Polynomials are naturally oscillatory functions. As the order of the polynomial increases, the number of oscillations increases and these oscillations become more severe. I'm simplifying here, the possibility of multiple and imaginary roots makes it a bit more complex, but the point is the same.

• Polynomials approach +/- infinity at a rate equal to the polynomial order as x goes to +/- infinity. This is often not a desired behavior.

• Computing polynomial coefficients for high order polynomials is typically an ill conditioned problem. This means that small errors (such as rounding in your computer) can create large changes in the answer. The linear system that must be solved involves a Vandermonde Matrix which can easily be ill conditioned.

I think that perhaps the heart of this issue is the distinction between curve fitting and interpolation.

Interpolation is used when you believe that your data is very accurate so you want your function to exactly match the data points. When you need values between your data points it's typically best to use a smooth function that matches the local trend of the data. Cubic or Hermite splines are often a good choice for this type of problem since they are much less sensitive to non-local (meaning at data points far away from a given point) changes or errors in the data and are less oscillatory than a polynomial. Consider the following data set:

x = 1   2   3   4   5   6   7   8   9  10
y = 1   1 1.1   1   1   1   1   1   1   1 A polynomial fit has much larger oscillations, especially near the edges of the data set, than a Hermite spline.

On the other hand, least squares approximation is a curve fitting technique. Curve fitting is used when you have some idea of the expected functionality of your data, but you don't need your function to exactly pass through all the data points. This is typical when the data may contain measurement errors or other imprecisions or when you wish to extract the general trend of the data. Least squares approximation is most often introduced in a course by using polynomials for the curve fitting because this results in a linear system that is relatively simple to solve using the techniques you likely learned earlier in your course. However, least squares techniques are much more general than just polynomial fits and can be used to fit any desired function to a data set. For example, if you expect an exponential growth trend in your data set, you can fit an exponential curve using least squares or linearized least squares.

Finally, choosing the correct function to fit your data is as important as correctly performing the interpolation or least squares computations. Doing so even allows the possibility of (cautious) extrapolation. Consider the following situation. Given population data (in millions of people) for the US from 2000-2010:

Year:  2000   2001   2002   2003   2004   2005   2006   2007   2008   2010
Pop.: 284.97 287.63 290.11 292.81 295.52 298.38 301.23 304.09 306.77 309.35

Using an exponential linearized least squares fit N(t)=A*exp(B*t) or a 10th order polynomial interpolant gives the following results: US population growth is not quite exponential, but I'll let you be the judge of the better fit.

• One point I would make with your graph of US population, from memory a good fit in the domain does not mean that it will extrapolate well. In that light, it might be misleading to show the large polynomial errors outside of the region where you have data. – Daryl Mar 26 '14 at 7:00
• @Daryl Agreed, that's why I emphasized that extrapolation should be done cautiously and that choosing an appropriate function is crucial in that case. – Doug Lipinski Mar 26 '14 at 13:11
• @DougLipinski Thanks for the insightful answer. Can you explain what you mean by linearized least squares in opposition to least squares ? – bela83 Apr 28 '16 at 12:15
• @bela83 Properly answering that would be too long for a comment. I think it would be a very good new question if you want to ask it. – Doug Lipinski Apr 28 '16 at 13:02
• @DougLipinski I'll give it a try ! – bela83 Apr 28 '16 at 13:16

Very unscientific, but a good rule of thumb is that 3rd degree polynomials are typically a good start, and in practice I have never seen anyone use more than a 6th degree one a get good results.

The "ideal" polynomial would be the lowest order one that represents the noiseless data well enough for your purpose.

If your data is small enough to allow it (it usually is), you can simply try fitting with higher order polynomials until you start seeing oscillations, which tend to be the sign of "overfitting".

An alternative would be a smoothing spline, but it very much depends on the application. Splines and smoothing splines are only good for interpolation. For noisy data I tend to prefer polynomial fits to smoothing splines, but the data I work with are typically well approximated by the polynomials.

A decently-effective approach that I've used with polynomial approximation is to calculate the least-squares polynomials for various degrees (e.g., from 1 to 10) and then choose the curve that minimizes the mean squared error at points halfway between consecutive (when sorted by x) points in your data set. This helps to rule out curves with too-severe oscillations.