# Maximum variance and useful information of dataset

I am reading through PCA and it says that the maximum variance principal component has most of the information. Can we apply that to any data set? If a data set has n attributes and most of the attributes show high variance then can we infer that the dataset has captured lot of useful information?

I am trying to understand how a high variance dataset contains useful information?

• What is PCA? By what metric is "most" defined? – Raphael Mar 5 '13 at 7:13
• Principal component analysis. It is used to reduce the dimensions or number of attributes by depicting the data as a linear combination of dependent attributes. So the linear combination of attributes(Principal components) which has the highest eigen value has the highest variance. Highest variance PC are considered over lower because they capture all the important information. – Sid Mar 5 '13 at 7:45
• So are you reading through a particular book? (In this context, was is "important" "information"?) – Raphael Mar 5 '13 at 10:23
• How do you quantify the information? By entropy? Or do you just want a simple explanation of why a PCA is useful all? If so, did you read the question on cross validated: stats.stackexchange.com/questions/tagged/pca ? – frafl Mar 5 '13 at 14:08
• You can't say "good is what PCA does" and ask "why does PCA do good?". You need to describe what "good" (or "information") means independently of PCA. – Raphael Mar 5 '13 at 17:18

## 1 Answer

I'll answer this, even thought the question is not well defined and would possibly be better suited for cross validated. But since it has some connection to information it's not totally off topic here.

Short version: Now, we can't apply that to any data set. Variance alone is not a good measure of information. PCA chooses the components with the highest eigenvalue, because they explain the input variance (i.e. euclidean distance from the origin). That's why you should normalize your input, if you don't want to emphasize a certain input component.

Long version:

Let's say you have two random variables $X$ and $Y:=10X$ where $E(X)=E(Y)=0$, $\mathrm{Var}(X)=1$ and thus $\mathrm{Var}(Y) = 100$. Now assume a third variable $K$ uniformly distributed among $\{1,\dots,k\}$ (this models drawing a random individual form your sample).

The variance before the PCA is not a sufficient measure for information, e.g. if we use $X$ and $Y$ to distinguish the individuals: $$P(K=k|X=x) = P(K=k|Y=10x)$$

The covariance matrix looks like this: $$\left(\matrix{1&10\\10 & 100}\right)$$

Using R's "eigen" we get:

$values  101 0$vectors
[,1]        [,2]
[1,] 0.09950372 -0.99503719
[2,] 0.99503719  0.09950372


If we now transform $X$ and $Y$ to $X'=M_{11}\cdot X + M_{21}\cdot Y$ and $Y'=M_{12}\cdot X + M_{22}\cdot Y$ (where $M$ is "vectors" from $R$), $X'$ contains "all the information": $$\forall y\;:P(K=k|X'=x) \neq P(K=k|Y'=y)=P(K=k)$$

But, of course, again $X''=10^{-100}\cdot X'$ contains "the same information".

• Good job on interpreting the question; I was inches away from closing the question. If you want, feel free to edit the question into shape! – Raphael Mar 6 '13 at 7:11