Why is the eigenvector in PCA taken to be unit norm? In deriving the eigenvectors for PCA, the vector is subject to the condition that it should be of unit length. Why is this so?
 A: The main aim of Principal Component Analysis (PCA) is to look for the directions on $\mathbb{R}^p$ that maximize the variance of the projected random vector $X=(X_1,\ldots,X_p)$. Specifically, the first PC can be defined as the unit vector $v_{(1)}\in\mathbb{R}^p$ such that 
$$v_{(1)}=\arg\max_{v\in\mathbb{R}^p,||v||=1}\mathbb{V}\mathrm{ar}\big[v^TX\big].$$
If you allow vectors that are not of unit norm in the maximization problem, then you will not get a proper solution, since variance of the projection can become arbitrarily large as long as the norm of the vector increases. For example, if $w=\lambda v$, with $v,w\in\mathbb{R}^p$ and $\lambda\to\infty$, then
$$\mathbb{V}\mathrm{ar}\big[w^TX\big]=\lambda^2\mathbb{V}\mathrm{ar}\big[v^TX\big]\to\infty\quad (\text{if }\mathbb{V}\mathrm{ar}\big[v^TX\big]\neq0).$$
This is the reason why you need an standardization of unit norm to constraint the search and avoid improper solutions.
A: It is not true that they "should be of unit length"; PCA works fine without using unit vectors given your data $x$ as long as you use a fixed arbitrary length $l$. 
Having said that you want to have the eigenvectors $\alpha_k$ of your covariance matrix $C$ to be unit vectors, ie. $\alpha_k^T \alpha_k = 1$, so you can: 


*

*Use the associated eigenvalue $\lambda_k$ as the variance of $\alpha_k^T x$.

*Use the eigenvectors as an axis of the ellipsoid fitted to $x$.


The first chapter from Jolliffe's Principal Component Analysis (Introduction) gives a more detailed (and nicer) exposition of these issues.
