John Detlefs Programmer

Principal Component Analysis

My next subject for bloggery is Principal Component Analysis (PCA) (its sibling Multidimensional scaling has been left out for a future post, but it is just as special, don’t worry). If I were to give a talk on PCA, the slides would be roughly ordered as follows:

  • A very short recap of dimension reduction
  • PCA, what it stands for, rough background, history
  • Eigenvectors (what are those?!)
  • Covariance (Because variance matrix didn’t sound cool enough)
  • The very fancy sounding method of Lagrange Multipliers (why they aren’t that hard)
  • Explain the PCA Algorithm
  • Random Walks: What are they, how are they taken on a configuration space
  • Interpreting the results after applying PCA on MD simulation data

In reality not going to follow these bullet points, if you want to get information pertaining to the first two points, please read some of my previous posts. The last two points are going to be a subject for a post next week.

Here are some good sources for those seeking to acquaint themselves with Linear Algebra and Statistics. Multivariate Statistics and PCA (Lessons two through 10) and the Feynman Lecture on Physics: Probability. The Feynman lectures on physics are so good and so accessible. Richard Feynman certainly had his flaws but teaching was not one of them. If you’re too busy to read those, here’s a quick summary of some important ideas I will be using.

What is a Linear Transformation, John?

Glad you asked, friend! Let’s just stick to linearity. for a function to be linear, it means that $f(a+b) = f(a) + f(b)$. As an example of a non-linear function, consider $ y = x^{3} $ . After plugging some numbers in we can see this is non-linear $ 2^{3} \neq 1^{3} + 1^{3} $.

A transformation at its most abstract is the description of an algorithm that gets an object A to become an object B. In linear algebra the transformation is being done on vectors belonging to a domain (where vectors exist before the transformation) space $V$ and a range (where vectors exist after the transformation) space $W$. For the purposes of our work, these are both $R^{n}$, the Cartesian product n-times of the real line. (The standard (x,y) coordinate system is the Cartesian product of the real line twice, or $R^{2}$)

When dealing with vector spaces, our linear transformation can be represented by a $m$-by-$n$ matrix, where $n$ is the dimension of the space we are sending a vector into (always less than or equal to m), and $m$ is the dimension of the vector space in which our original vector (or set of vectors) being transformed originally exists in. So if we have some set of $k$ vectors being transformed, the matrices will be have row-by-column sizes: . These maps can be scalings, rotations, shearings and more.

What is a vector, what is an eigenvector?

Good question! A vector has a magnitude (which is just some positive number for anything that we are doing) and a direction (a property that is drawn from the vector space to which the vector belongs). Being told to walk 10 paces due north is to follow a vector with magnitude 10 and direction north. Vectors are presented as if they are centered at the origin, and their head is reflects their magnitude and direction. This allows some consistency when discussing things, when we are given the vector $(1,1)$ in $R^2$ we know it is centered at the origin, and thus has magnitude (from the distance formula) of $\sqrt{2}$ and the direction is 45 degrees from the horizontal axis. An eigenvector sounds a lot scarier than it is; the purpose of an eigenvector is to answer the question, ‘what vectors don’t change direction under a given linear transformation?’ mona lisa eigenvector

This picture is stolen from wikipedia, but it should be clear that the blue vector is an eigenvector of this transformation, while the red vector is not.

The standard equation given when an eigenvalue problem is posed is:

$M$ is some linear transformation, $v$ is an eigenvector we are trying to find, and $\lambda$ is the corresponding eigenvalue.

From this equation, we can see that eigenvector-eigenvalue pairs are not unique; direction is not a unique property of a vector. If we find a vector $v$ that satisfies this equation for our linear transformation $M$, scaling the vector by some constant $\alpha$ will simply change the eigenvalue associated with the solution. The vector $(1,1)$ in $R^2$ has the same direction as $(2,2)$ in $R^2$. If one of these vectors isn’t subject to a direction change (and therefore an eigenvector), then the other must be as well, because the eigenvector-ness (yes, I just coined this phrase) applies to all vectors with the same direction.

For those of you more familiar with Linear Algebra, this should not be confused with the fact that a linear transformation can have degenerate eigenvalues. This concept of degenerate eigenvalues comes up when the rank of the matrix representation of a linear transformation is less than its dimension, but given that our transformation has rank equal to to the dimension of the vector space, we can ignore this.

Statistics in ~300 words

Sticking one dimension, plenty of data seems to be random while also favoring a central point. Consider the usual example of height across a population of people. Height can be thought of as a random variable. But this isn’t random in the way that people might think about randomness without some knowledge of statistics. There is a central point where the heights of people tend to cluster, and the likelihood of someone being taller or shorter than this central point decreases on a ‘bell-curve’.

normal

This is called a normal distribution. Many datasets can be thought of as describing the a set of outcomes for some random variable in nature. These outcomes are distributed in some fashion. In our example, the mean of the data is the average height over the entire population. The variance of our data is how far the height of a is spread out from its mean. When the distribution of outcomes follows a bell-curve such as it does in our example, the distribution is referred to as normal. (There are some more technical details, but the phrase normal stems from the fact that total area under the bell-curve defined by the normal distribution is equal to one.) When the data we want to describe is reflects more than one random variable this is a multivariate distribution. multivariate Statistics introduces the concept of covariance to describe the relationship that random variables have with one another. The magnitude of a covariance indicates in some fashion the relationship between random variables; it is not easily interpretable without some set of constraints on the covariances, which will come up in Principal Component Analysis. The sign of the covariance between two random variables (X,Y) indicates if the two points are inversely related or directly related. A negative covariance between X and Y means that increasing X decreases while a positive covariance means that increasing X increases Y. The normalized version of covariance is called the correlation coefficient and might be more familiar to those previously acquainted with statistics.

Constrained Optimization

Again, please do look at the Feynman Lectures, if you’re unfamiliar with statistics look at the Penn State material for the ideas I just went over to better understand them. The last subject I want to broach before getting into the details of PCA is optimization with Lagrange Multipliers.

Lagrange Multipliers are one method of solving a constrained optimization problem. Such a problem requires an objective function to be optimized and constraints to optimize against. An objective function is any function that we wish to maximize or minimize to reflect some target quantity achieving an optimum value. In short, the method of Lagrange Multipliers creates a system of linear equations such that we can solve for a term $\lambda$ that shows when the objective function achieves a maximum subject to constraints. In the case the of PCA, the function we want to maximize is $M^{T} cov(X) M$, describing the covariance of our data matrix $X$.

PCA

PCA-vectors

Although it is not quite a fortuitous circumstance, the principal components of the covariance matrix are precisely it’s eigenvectors. For a multivariate dataset generated by $n$ random variables, PCA will return a sequence of $n$ eigenvectors each describing more covariance in the dataset than the next. The picture above presents an example of the eigenvectors reflecting the covariance of a 2-dimensional multivariate dataset.

To be perfectly honest, I don’t know a satisfying way to explain why the eigenvectors are the principal components. The best explanation I can come up with is that the algorithm for PCA is in correspondence with the algorithm for eigenvalue decomposition. It’s one of those things where I should be able to provide a proof for why the two problems are the same, but I cannot at the moment. (Commence hand-waving…)

Let’s look at the algorithm for Principal Component Analysis to better understand things. PCA is an iterative method that seeks to create a sequence of vectors that describe the covariance in a collection of data, each vector describing more than those that will follow. This introduces an optimization problem, an issue that I referenced earlier. In order to guarantee uniqueness of our solution, this optimization is subject to constraints using Lagrange Multipliers. remember, the video provides an example of a constrained optimization problem from calculus. Principal Component Analysis finds this set of vectors by creating a linear transformation $M$ that maximizes the covariance objective function given below.

PCA’s objective function is $trace(M^{T} cov(X) M)$, we are looking to maximize this. From the Penn State lectures:

Earlier in the course we defined the total variation of X as the trace of the variance-covariance matrix, or if you like, the sum of the variances of the individual variables. This is also equal to the sum of the eigenvalues.

In the first step of PCA, we can think of our objective function as:

$a$ in this case is a single vector

We seek to maximize this a term such that the sum of the squares of the coefficients of a is equal to one. (In math terms this is saying that the $L^2$ norm is one). This constraint is introduced to ensure that a unique answer is obtained, (remember eigenvectors are not unique, this is the same process one would undertake to get a unique sequence of eigenvectors from a decomposition!). In the second step, the objective function is :

$B$ consists of $a$ and a new vector $b$

We look to maximize $b$ such that it explains covariance not previously explained by the first, for this reason we introduce an optimization problem with two constraints:

  • $a$ remains the same
  • The sum of the squares of the coefficients of $b$ equals one
  • None of the covariance explained by vector $a$ is explained by $b$, (the two vectors are orthogonal, another property of eigenvectors!)

This process is repeated for the entire transformation $M$. This gives us a sequence of eigenvalues that each reflect some fraction of the covariance, and sum to one. For our previously mentioned n-dimensional multivariate dataset, generally some number $k \lt n$ of eigenvectors explain the total variance well enough to ignore the $n - k$ vectors remaining. This gives a set of principal components to investigate.

It might follow that this corresponds to an eigenvalue problem:

If it doesn’t appear to be clear, let’s step back and look again at eigenvectors. As I said earlier, eigenvectors provide insight into what vectors don’t change direction under a linear map. We are trying find a linearly independent set of vectors that provide insight into the structure of our covariance from our multivariate distribution. Eigenvectors are either the same or orthogonal. This algorithm we described, is precisely the algorithm one would use to find the eigenvectors of any (full-rank) linear transformation, not just the linear transformation done in Principal Component Analysis.

Chemistry application and interpretation:

Before PCA is done on an MD Simulation, we have to consider what goals we have for the analysis of results. We are searching to arrange the data outputted by PCA such it gives us intuition into some physical behavior of our system. This is usually done a single structure, and in order to focus insight on relative structural change rather than some sort of translational motion of the entire structure, an alignment of an entire trajectory to some structure of interest by minimization of the Root Mean Square Distance must be done prior to analysis. After RMS alignment, any variance in data should be variance due to changes in structure.

Cecilia Clementi has this quote in her Free Energy Landscapes paper:

Essentially, PCA computes a hyperplane that passes through the data points as best as possible in a least-squares sense. The principal components are the tangent vectors that describe this hyperplane

How do we interpret n-dimensional physical data from the tangent vectors of an k-dimensional hyperplane embedded in this higher dimensional space? This is all very mathematical and abstract, . What we do is reduce the analysis to some visually interpretable subset of components, and see if there is any indication of clustering that occurs.

Remember, we have an explicit linear map relating the higher dimensional space to the lower-dimensional space . By taking our trajectory and projecting it onto one of the set of eigenvector components of our analysis, we can extract embeddings in different ways, from Clementi again:

So, the first principal component corresponds to the best possible projection onto a line, the first two correspond to the best possible projection onto a plane, and so on. Clearly, if the manifold of interest is inherently non-linear the low-dimensional e mbedding obtained by means of PCA is severely distorted… The fact that empirical reaction coordinates routinely used in protein folding studies can not be reduced to a linear combination of the Cartesian coordinates underscores the inadequacy of linear dimensionality reduction techniques to characterize a folding landscape.

Again, this is all esoteric for the lay reader. What is a manifold, a reaction coordinate, a linear combination of cartesian coordinates? All we should know is that PCA is a limited investigational tool for complex systems, the variance the principal components explain should not necessarily be interpreted as physical parameters governing the behavior of a system.

My mentor max has a great Jupyter notebook up demonstrating PCA done on MD simulations here. All of these topics are covered in the notebook and should be relatively accessible if you understand what I’ve said so far. In my next post I will write about how I will be implementing PCA as a module in MDAnalysis.

-John