A geometric interpretation of the covariance

matrix
Contents [hide] [hide]

1 Introduction

2 Eigendecomposition of a covariance matrix

3 Covariance matrix as a linear transformation

4 Conclusion

Introduction
In this article, we provide an intuitive, geometric interpretation of the covariance matrix, by
exploring the relation between linear transformations and the resulting data covariance. Most
textbooks explain the shape of data based on the concept of covariance matrices. Instead, we
take a backwards approach and explain the concept of covariance matrices based on the shape of
data.
In a previous article, we discussed the concept of variance, and provided a derivation and proof of
the well known formula to estimate the sample variance. Figure 1 was used in this article to show
that the standard deviation, as the square root of the variance, provides a measure of how much
the data is spread across the feature space.

Figure 1. Gaussian density function. For normally distributed data, 68% of the samples fall within
the interval defined by the mean plus and minus the standard deviation.
We showed that an unbiased estimator of the sample variance can be obtained by:

(1)
However, variance can only be used to explain the spread of the data in the directions parallel to
the axes of the feature space. Consider the 2D feature space shown by figure 2:

Figure 2. The diagnoal spread of the data is captured by the covariance.

For this data, we could calculate the variance
variance

in the x-direction and the

in the y-direction. However, the horizontal spread and the vertical spread of the

data does not explain the clear diagonal correlation. Figure 2 clearly shows that on average, if the
x-value of a data point increases, then also the y-value increases, resulting in a positive correlation.
This correlation can be captured by extending the notion of variance to what is called the
covariance of the data:

(2)
For 2D data, we thus obtain

and

. These four values can be

summarized in a matrix, called the covariance matrix:

(3)
If x is positively correlated with y, y is also positively correlated with x. In other words, we can state
that

. Therefore, the covariance matrix is always a symmetric matrix with the

variances on its diagonal and the covariances off-diagonal. Two-dimensional normally distributed

data is explained completely by its mean and its

a
a

covariance matrix. Similarly,

covariance matrix is used to capture the spread of three-dimensional data, and

covariance matrix captures the spread of N-dimensional data.

Figure 3 illustrates how the overall shape of the data defines the covariance matrix:

Figure 3. The covariance matrix defines the shape of the data. Diagonal spread is captured by the
covariance, while axis-aligned spread is captured by the variance.

Eigendecomposition of a covariance matrix

In the next section, we will discuss how the covariance matrix can be interpreted as a linear
operator that transforms white data into the data we observed. However, before diving into the
technical details, it is important to gain an intuitive understanding of how eigenvectors and
eigenvalues uniquely define the covariance matrix, and therefore the shape of our data.
As we saw in figure 3, the covariance matrix defines both the spread (variance), and the
orientation (covariance) of our data. So, if we would like to represent the covariance matrix with a

vector and its magnitude, we should simply try to find the vector that points into the direction of the
largest spread of the data, and whose magnitude equals the spread (variance) in this direction.
If we define this vector as
as
vector

, then the projection of our data

, and the variance of the projected data is

onto this vector is obtained

. Since we are looking for the

that points into the direction of the largest variance, we should choose its components

such that the covariance matrix

any function of the form

of the projected data is as large as possible. Maximizing

with respect to

, where

is a normalized unit vector, can be

formulated as a so called Rayleigh Quotient. The maximum of such a Rayleigh Quotient is

obtained by setting

equal to the largest eigenvector of matrix

In other words, the largest eigenvector of the covariance matrix always points into the direction of
the largest variance of the data, and the magnitude of this vector equals the corresponding
eigenvalue. The second largest eigenvector is always orthogonal to the largest eigenvector, and
points into the direction of the second largest spread of the data.
Now lets have a look at some examples. In an earlier article we saw that a linear transformation
matrix

is completely defined by its eigenvectors and eigenvalues. Applied to the covariance

matrix, this means that:

(4)
where

is an eigenvector of

, and

is the corresponding eigenvalue.

If the covariance matrix of our data is a diagonal matrix, such that the covariances are zero, then
this means that the variances must be equal to the eigenvalues

. This is illustrated by figure 4,

where the eigenvectors are shown in green and magenta, and where the eigenvalues clearly
equal the variance components of the covariance matrix.

Figure 4. Eigenvectors of a covariance matrix

However, if the covariance matrix is not diagonal, such that the covariances are not zero, then the
situation is a little more complicated. The eigenvalues still represent the variance magnitude in the
direction of the largest spread of the data, and the variance components of the covariance matrix
still represent the variance magnitude in the direction of the x-axis and y-axis. But since the data is
not axis aligned, these values are not the same anymore as shown by figure 5.

Figure 5. Eigenvalues versus variance

By comparing figure 5 with figure 4, it becomes clear that the eigenvalues represent the variance
of the data along the eigenvector directions, whereas the variance components of the covariance
matrix represent the spread along the axes. If there are no covariances, then both values are
equal.

Covariance matrix as a linear transformation

Now lets forget about covariance matrices for a moment. Each of the examples in figure 3 can
simply be considered to be a linearly transformed instance of figure 6:

Figure 6. Data with unit covariance matrix is called white data.

Let the data shown by figure 6 be
obtained by linearly transforming

, then each of the examples shown by figure 3 can be

:

(5)
where

is a transformation matrix consisting of a rotation matrix

(6)
These matrices are defined as:

(7)
where

is the rotation angle, and:

and a scaling matrix

(8)
where

and

are the scaling factors in the x direction and the y direction respectively.

In the following paragraphs, we will discuss the relation between the covariance matrix
linear transformation matrix

, and the

Lets start with unscaled (scale equals 1) and unrotated data. In statistics this is often refered to as
white data because its samples are drawn from a standard normal distribution and therefore
correspond to white (uncorrelated) noise:

Figure 7. White data is data with a unit covariance matrix.

The covariance matrix of this white data equals the identity matrix, such that the variances and
standard deviations equal 1 and the covariance equals zero:

(9)

Now lets scale the data in the x-direction with a factor 4:

(10)
The data

now looks as follows:

Figure 8. Variance in the x-direction results in a horizontal scaling.

The covariance matrix

of

is now:

(11)
Thus, the covariance matrix
transformation

(12)

of the resulting data

is related to the linear

that is applied to the original data as follows:

, where

However, although equation (12) holds when the data is scaled in the x and y direction, the
question rises if it also holds when a rotation is applied. To investigate the relation between the
linear transformation matrix

and the covariance matrix

in the general case, we will therefore

try to decompose the covariance matrix into the product of rotation and scaling matrices.
As we saw earlier, we can represent the covariance matrix by its eigenvectors and eigenvalues:

(13)
where

is an eigenvector of

, and

is the corresponding eigenvalue.

Equation (13) holds for each eigenvector-eigenvalue pair of matrix

. In the 2D case, we obtain

two eigenvectors and two eigenvalues. The system of two equations defined by equation (13) can
be represented efficiently using matrix notation:

(14)
where

is the matrix whose columns are the eigenvectors of

and

is the diagonal matrix

whose non-zero elements are the corresponding eigenvalues.

This means that we can represent the covariance matrix as a function of its eigenvectors and
eigenvalues:

(15)
Equation (15) is called the eigendecomposition of the covariance matrix and can be obtained
using aSingular Value Decomposition algorithm. Whereas the eigenvectors represent the
directions of the largest variance of the data, the eigenvalues represent the magnitude of this
variance in those directions. In other words,

represents a rotation matrix, while

a scaling matrix. The covariance matrix can thus be decomposed further as:

(16)
where

is a rotation matrix and

is a scaling matrix.

represents

In equation (6) we defined a linear transformation

matrix,

. Furthermore, since

Therefore,

. Since

is an orthogonal matrix,

is a diagonal scaling
.

. The covariance matrix can thus be written as:

(17)
In other words, if we apply the linear transformation defined by
data

shown by figure 7, we obtain the rotated and scaled data

matrix

to the original white

with covariance

. This is illustrated by figure 10:

Figure 10. The covariance matrix represents a linear transformation of the original data.
The colored arrows in figure 10 represent the eigenvectors. The largest eigenvector, i.e. the
eigenvector with the largest corresponding eigenvalue, always points in the direction of the largest
variance of the data and thereby defines its orientation. Subsequent eigenvectors are always
orthogonal to the largest eigenvector due to the orthogonality of rotation matrices.

Conclusion
In this article we showed that the covariance matrix of observed data is directly related to a linear
transformation of white, uncorrelated data. This linear transformation is completely defined by the
eigenvectors and eigenvalues of the data. While the eigenvectors represent the rotation matrix,
the eigenvalues correspond to the square of the scaling factor in each dimension.

How to draw a covariance error ellipse?

Contents [hide] [hide]

1 Introduction
2 Axis-aligned confidence ellipses
3 Arbitrary confidence ellipses
4 Source Code
5 Conclusion

Introduction
In this post, I will show how to draw an error ellipse, a.k.a. confidence ellipse, for 2D
normally distributed data. The error ellipse represents an iso-contour of the Gaussian
distribution, and allows you to visualize a 2D confidence interval. The following figure
shows a 95% confidence ellipse for a set of 2D normally distributed data samples. This
confidence ellipse defines the region that contains 95% of all samples that can be drawn
from the underlying Gaussian distribution.

Figure 1. 2D confidence ellipse for normally distributed data

In the next sections we will discuss how to obtain confidence ellipses for different
confidence values (e.g. 99% confidence interval), and we will show how to plot these
ellipses using Matlab or C++ code.

Axis-aligned confidence ellipses

Before deriving a general methodology to obtain an error ellipse, lets have a look at the
special case where the major axis of the ellipse is aligned with the X-axis, as shown by the
following figure:

Figure 2. Confidence ellipse for uncorrelated Gaussian data

The above figure illustrates that the angle of the ellipse is determined by the covariance of
the data. In this case, the covariance is zero, such that the data is uncorrelated, resulting in
an axis-aligned error ellipse.

Table 1. Covariance matrix of the data shown in Figure 2

8.4213

0.9387

Furthermore, it is clear that the magnitudes of the ellipse axes depend on the variance of
the data. In our case, the largest variance is in the direction of the X-axis, whereas the
smallest variance lies in the direction of the Y-axis.
In general, the equation of an axis-aligned ellipse with a major axis of length
and a
minor axis of length , centered at the origin, is defined by the following equation:

(1)
In our case, the length of the axes are defined by the standard deviations
the data such that the equation of the error ellipse becomes:

(2)

and

of

where defines the scale of the ellipse and could be any arbitrary number (e.g. s=1). The
question is now how to choose , such that the scale of the resulting ellipse represents a
chosen confidence level (e.g. a 95% confidence level corresponds to s=5.991).
Our 2D data is sampled from a multivariate Gaussian with zero covariance. This means
that both the x-values and the y-values are normally distributed too. Therefore, the left
hand side of equation (2) actually represents the sum of squares of independent normally
distributed data samples. The sum of squared Gaussian data points is known to be
distributed according to a so called Chi-Square distribution. A Chi-Square distribution is
defined in terms of degrees of freedom, which represent the number of unknowns. In our
case there are two unknowns, and therefore two degrees of freedom.
Therefore, we can easily obtain the probability that the above sum, and thus equals a
specific value by calculating the Chi-Square likelihood. In fact, since we are interested in a
confidence interval, we are looking for the probability that is less then or equal to a
specific value which can easily be obtained using the cumulative Chi-Square distribution.
As statisticians are lazy people, we usually dont try to calculate this probability, but simply
look it up in a probability table: https://people.richland.edu/james/lecture/m170/tbl-chi.html.
For example, using this probability table we can easily find that, in the 2-degrees of
freedom case:

Therefore, a 95% confidence interval corresponds to s=5.991. In other words, 95% of the
data will fall inside the ellipse defined as:

(3)
Similarly, a 99% confidence interval corresponds to s=9.210 and a 90% confidence interval
corresponds to s=4.605.
The error ellipse show by figure 2 can therefore be drawn as an ellipse with a major axis
length equal to

and the minor axis length to

Arbitrary confidence ellipses

In cases where the data is not uncorrelated, such that a covariance exists, the resulting
error ellipse will not be axis aligned. In this case, the reasoning of the above paragraph
only holds if we temporarily define a new coordinate system such that the ellipse becomes
axis-aligned, and then rotate the resulting ellipse afterwards.
In other words, whereas we calculated the variances
and
parallel to the x-axis and
y-axis earlier, we now need to calculate these variances parallel to what will become the
major and minor axis of the confidence ellipse. The directions in which these variances
need to be calculated are illustrated by a pink and a green arrow in figure 1.

Figure 1. 2D confidence ellipse for normally distributed data

These directions are actually the directions in which the data varies the most, and are
defined by the covariance matrix. The covariance matrix can be considered as a matrix
that linearly transformed some original data to obtain the currently observed data. In a
previous article about eigenvectors and eigenvalues we showed that the direction vectors
along such a linear transformation are the eigenvectors of the transformation matrix.
Indeed, the vectors shown by pink and green arrows in figure 1, are the eigenvectors of the
covariance matrix of the data, whereas the length of the vectors corresponds to the
eigenvalues.
The eigenvalues therefore represent the spread of the data in the direction of the
eigenvectors. In other words, the eigenvalues represent the variance of the data in the
direction of the eigenvectors. In the case of axis aligned error ellipses, i.e. when the
covariance equals zero, the eigenvalues equal the variances of the covariance matrix and
the eigenvectors are equal to the definition of the x-axis and y-axis. In the case of arbitrary
correlated data, the eigenvectors represent the direction of the largest spread of the data,
whereas the eigenvalues define how large this spread really is.
Thus, the 95% confidence ellipse can be defined similarly to the axis-aligned case, with the
major axis of length
and the minor axis of length
,
where
and
represent the eigenvalues of the covariance matrix.
To obtain the orientation of the ellipse, we simply calculate the angle of the largest
eigenvector towards the x-axis:

(4)

where
is the eigenvector of the covariance matrix that corresponds to the largest
eigenvalue.
Based on the minor and major axis lengths and the angle between the major axis and
the x-axis, it becomes trivial to plot the confidence ellipse. Figure 3 shows error ellipses for
several confidence values:

Confidence ellipses for normally distributed data

Source Code
Matlab source code
C++ source code (uses OpenCV)

Conclusion
In this article we showed how to obtain the error ellipse for 2D normally distributed data,
according to a chosen confidence value. This is often useful when visualizing or analyzing
data and will be of interest in a future article about PCA.

Why divide the sample variance by N-1?

Contents [hide] [hide]

1 Introduction
2 Minimum variance, unbiased estimators
o 2.1 Parameter bias
o 2.2 Parameter variance
3 Maximum Likelihood estimation
4 Estimating the variance if the mean is known
o 4.1 Parameter estimation
o 4.2 Performance evaluation
5 Estimating the variance if the mean is unknown
o 5.1 Parameter estimation
o 5.2 Performance evaluation
o 5.3 Fixing the bias
6 Conclusion

Introduction
In this article, we will derive the well known formulas for calculating the
mean and the variance of normally distributed data, in order to answer the
question in the articles title. However, for readers who are not interested in
the why of this question but only in the when, the answer is quite simple:
If you have to estimate both the mean and the variance of the data (which is
typically the case), then divide by N-1, such that the variance is obtained as:

If, on the other hand, the mean of the true population is known such that only
the variance needs to be estimated, then divide by N, such that the variance is
obtained as:

Whereas the former is what you will typically need, an example of the latter
would be the estimation of the spread of white Gaussian noise. Since the mean
of white Gaussian noise is known to be zero, only the variance needs to be
estimated in this case.

If data is normally distributed we can completely characterize it by its mean and its
variance
. The variance is the square of the standard deviation which represents the
average deviation of each data point to the mean. In other words, the variance represents
the spread of the data. For normally distributed data, 68.3% of the observations will have a
value between
and
. This is illustrated by the following figure which shows a
Gaussian density function with mean
and variance
:

Figure 1. Gaussian density function. For normally distributed data, 68% of the samples fall
within the interval defined by the mean plus and minus the standard deviation.
Usually we do not have access to the complete population of the data. In the above
example, we would typically have a few observations at our disposal but we do not have
access to all possible observations that define the x-axis of the plot. For example, we might
have the following set of observations:

Table 1

Observation ID
Observation 1
Observation 2
Observation 3
Observation 4
Observation 5

Observed Value
10
12
7
5
11

If we now calculate the empirical mean by summing up all values and dividing by the
number of observations, we have:

(1)
Usually we assume that the empirical mean is close to the actually unknown mean of the
distribution, and thus assume that the observed data is sampled from a Gaussian
distribution with mean
. In this example, the actual mean of the distribution is 10, so
the empirical mean indeed is close to the actual mean.
The variance of the data is calculated as follows:

(2
Again, we usually assume that this empirical variance is close to the real and unknown
variance of underlying distribution. In this example, the real variance was 9, so indeed the
empirical variance is close to the real variance.
The question at hand is now why the formulas used to calculate the empirical mean and
the empirical variance are correct. In fact, another often used formula to calculate the
variance, is defined as follows:

(3)
The only difference between equation (2) and (3) is that the former divides by N-1,
whereas the latter divides by N. Both formulas are actually correct, but when to use which
one depends on the situation.
In the following sections, we will completely derive the formulas that best approximate the
unknown variance and mean of a normal distribution, given a few samples from this
distribution. We will show in which cases to divide the variance by N and in which cases to
normalize by N-1.
A formula that approximates a parameter (mean or variance) is called an estimator. In the
following, we will denote the real and unknown parameters of the distribution by and
.
The estimators, e.g. the empirical average and empirical variance, are denoted
as and
.
To find the optimal estimators, we first need an analytical expression for the likelihood of
observing a specific data point , given the fact that the population is normally distributed
with a given mean and standard deviation . A normal distribution with known
parameters is usually denoted as

. The likelihood function is then:

(4)
To calculate the mean and variance, we obviously need more than one sample from this
distribution. In the following, let vector
be a vector that contains all
the available samples (e.g. all the values from the example in table 1). If all these samples
are statistically independent, we can write their joint likelihood function as the sum of all
individual likelihoods:

(5)
Plugging equation (4) into equation (5) then yields an analytical expression for this joint
probability density function:

(6)
Equation (6) will be important in the next sections and will be used to derive the well known
expressions for the estimators of the mean and the variance of a Gaussian distribution.

Minimum variance, unbiased estimators

To determine if an estimator is a good estimator, we first need to define what a good
estimator really is. The goodness of an estimator depends on two measures, namely its
bias and its variance (yes, we will talk about the variance of the mean-estimator and the
variance of the variance-estimator). Both measures are briefly discussed in this section.

Parameter bias
Imagine that we could obtain different (disjoint) subsets of the complete population. In
analogy to our previous example, imagine that, apart from the data in Table 1, we also
have a Table 2 and a Table 3 with different observations. Then a good estimator for the
mean, would be an estimator that on average would be equal to the real mean. Although
we can live with the idea that the empirical mean from one subset of data is not equal to
the real mean like in our example, a good estimator should make sure that the average of

the estimated means from all subsets is equal to the real mean. This constraint is
expressed mathematically by stating that the Expected Value of the estimator should equal
the real parameter value:

(7)
If the above conditions hold, then the estimators are called unbiased estimators. If the
conditions do not hold, the estimators are said to be biased, since on average they will
either underestimate or overestimate the true value of the parameter.

Parameter variance
Unbiased estimators guarantee that on average they yield an estimate that equals the real
parameter. However, this does not mean that each estimate is a good estimate. For
instance, if the real mean is 10, an unbiased estimator could estimate the mean as 50 on
one population subset and as -30 on another subset. The expected value of the estimate
would then indeed be 10, which equals the real parameter, but the quality of the estimator
clearly also depends on the spread of each estimate. An estimator that yields the
estimates (10, 15, 5, 12, 8) for five different subsets of the population is unbiased just like
an estimator that yields the estimates (50, -30, 100, -90, 10). However, all estimates from
the first estimator are closer to the true value than those from the second estimator.
Therefore, a good estimator not only has a low bias, but also yields a low variance. This
variance is expressed as the mean squared error of the estimator:

A good estimator is therefore is a low bias, low variance estimator. The optimal estimator,
if such estimator exists, is then the one that has no bias and a variance that is lower than
any other possible estimator. Such an estimator is called the minimum variance, unbiased
(MVU) estimator. In the next section, we will derive the analytical expressions for the mean
and the variance estimators of a Gaussian distribution. We will show that the MVU
estimator for the variance of a normal distribution requires us to divide the variance by
under certain assumptions, and requires us to divide by N-1 if these assumptions do not
hold.

Maximum Likelihood estimation

Although numerous techniques can be used to obtain an estimator of the parameters
based on a subset of the population data, the simplest of all is probably the maximum
likelihoodapproach.

The probability of observing

was defined by equation (6) as
. If we fix
and
in this function, while letting
vary, we obtain the Gaussian distribution as plotted
by figure 1. However, we could also choose a fixed
and let and/or vary. For
example, we can choose
like in our previous example. We also
choose a fixed
, and we let
vary. Figure 2 shows the plot of each different
value of
for the distribution with the proposed fixed
and :

Figure 2. This plot shows the likelihood of observing fixed data

if the data is normally
distributed with a chosen, fixed
, plotted against various values of a varying
.
In the above figure, we calculated the likelihood
by varying
for a
fixed
. Each point in the resulting curve represents the likelihood that
observation is a sample from a Gaussian distribution with parameter
. The parameter
value that corresponds to the highest likelihood is then most likely the parameter that
defines the distribution our data originated from. Therefore, we can determine the
optimal
by finding the maximum in this likelihood curve. In this example, the maximum
is at
, such that the standard deviation is
calculate the variance in the traditional way, with a given
equal to 7.8:

. Indeed if we would
, we would find that it is

Therefore, the formula to compute the variance based on the sample data is simply
derived by finding the peak of the maximum likelihood function. Furthermore, instead of

fixing , we let both and

vary at the same time. Finding both estimators then
corresponds to finding the maximum in a two-dimensional likelihood function.
To find the maximum of a function, we simply set its derivative to zero. If we want to find
the maximum of a function with two variables, we calculate the partial derivative towards
each of these variables and set both to zero. In the following, let
be the optimal
estimator for the population mean as obtained using the maximum likelihood method, and
let
be the optimal estimator for the variance. To maximize the likelihood function we
simply calculate its (partial) derivatives and set them to zero as follows:

and

In the following paragraphs we will use this technique to obtain the MVU estimators of both
and . We consider two cases:
The first case assumes that the true mean of the distribution is known. Therefore, we
only need to estimate the variance and the problem then corresponds to finding the
maximum in a one-dimensional likelihood function, parameterized by
. Although this
situation does not occur often in practice, it definitely has practical applications. For
instance, if we know that a signal (e.g. the color value of a pixel in an image) should have a
specific value, but the signal has been polluted by white noise (Gaussian noise with zero
mean), then the mean of the distribution is known and we only need to estimate the
variance.
The second case deals with the situation where both the true mean and the true variance
are unknown. This is the case you would encounter most and where you would obtain an
estimate of the mean and the variance based on your sample data.
In the next paragraphs we will show that each case results in a different MVU estimator.
More specific, the first case requires the variance estimator to be normalized by
to be
MVU, whereas the second case requires division by
to be MVU.

Estimating the variance if the mean is known

Parameter estimation
If the true mean of the distribution is known, then the likelihood function is only
parameterized on
. Obtaining the maximum likelihood estimator then corresponds to
solving:

(8)
However, calculating the derivative of
, defined by equation (6) is rather involved
due to the exponent in the function. In fact, it is much easier to maximize the log-likelihood
function instead of the likelihood function itself. Since the logarithm is a monotonous
function, the maximum will be the same. Therefore, we solve the following problem
instead:

(9)
In the following we set
to obtain a simpler notation. To find the maximum of the
log-likelihood function, we simply calculate the derivative of the logarithm of equation (6)
and set it to zero:

It is clear that if

, then the only possible solution to the above is:

(10)
Note that this maximum likelihood estimator for

is indeed the traditional formula to

calculate the variance of normal data. The normalization factor is

However, the maximum likelihood method does not guarantee to deliver an unbiased
estimator. On the other hand, if the obtained estimator is unbiased, then the maximum
likelihood method does guarantee that the estimator is also minimum variance and thus
MVU. Therefore, we need to check if the estimator in equation (10) is unbiassed.

Performance evaluation
To check if the estimator defined by equation (10) is unbiassed, we need to check if the
condition of equation (7) holds, and thus if

To do this, we plug equation (10) into

and write:

Furthermore, an important property of variance is that the true variance

as
such that
in the above equation yields:

can be written

. Using this property

Since
, the condition shown by equation (7) holds, and therefore the obtained
estimator for the variance of the data is unbiassed. Furthermore, because the maximum
likelihood method guarantees that an unbiased estimator is also minimum variance (MVU),
this means that no other estimator exists that can do better than the one obtained here.
Therefore, we have to divide by
instead of
while calculating the variance of
normally distributed data, if the true mean of the underlying distribution is known.

Estimating the variance if the mean is unknown

Parameter estimation
In the previous section, the true mean of the distribution was known, such that we only had
to find an estimator for the variance of the data. However, if the true mean is not known,
then an estimator has to be found for the mean too. Furthermore, this mean estimate is
used by the variance estimator. As a result, we will show that the earlier obtained estimator
for the variance is no longer unbiassed. Furthermore, we will show that we can unbias the
estimator in this case by dividing by
instead of by , which slightly increases the
variance of the estimator.
As before, we use the maximum likelihood method to obtain the estimators based on the
log-likelihood function. We first find the ML estimator for :

If

, then it is clear that the above equation only has a solution if:

(11)
Note that indeed this is the well known formula to calculate the mean of a distribution.
Although we all knew this formula, we now proved that it is the maximum likelihood
estimator for the true and unknown mean of a normal distribution. For now, we will just
assume that the estimator that we found earlier for the variance , defined by equation (10),
is still the MVU variance estimator. In the next section however, we will show that this
estimator is no longer unbiased now.

Performance evaluation
To check if the estimator for the true mean
the condition of equation (7) holds:

is unbiassed, we have to make sure that

Since
, this means that the obtained estimator for the mean of the distribution is
unbiassed. Since the maximum likelihood method guarantees to deliver the minimum
variance estimator if the estimator is unbiassed, we proved that is the MVU estimator of
the mean.
To check if the earlier found estimator for the variance is still unbiassed if it is based on
the empirical mean instead of the true mean , we simply plug the obtained
estimator into the earlier derived estimator of equation

(10):
To check if the estimator is still unbiased, we now need to check again if the condition of
equation (7) holds:

As mentioned in the previous section, an important property of variance is that the true
variance
that

can be written as

such
. Using this property in the above equation yields:

Since clearly
, this shows that estimator for the variance of the distribution is no
longer unbiassed. In fact, this estimator on average underestimates the true variance with
a factor
. As the number of samples approaches infinity (
), this bias
converges to zero. For small sample sets however, the bias is signification and should be
eliminated.

Fixing the bias

Since the bias is merely a factor, we can eliminate it by scaling the biased estimator
defined by equation (10) by the inverse of the bias. We therefore define a new, unbiased

estimate

as follows:

This estimator is now unbiassed and indeed resembles the traditional formula to calculate
the variance, where we divide by
instead of . However, note that the resulting
estimator is no longer the minimum variance estimator, but it is the estimator with the
minimum variance amongst all unbiased estimators. If we divide by , then the estimator
is biassed, and if we divide by
, the estimator is not the minimum variance estimator.
However, in general having a biased estimator is much worse than having a slightly higher
variance estimator. Therefore, if the mean of the population is unknown, division
by
should be used instead of division by .

Conclusion
In this article, we showed where the usual formulas for calculating the mean and the
variance of normally distributed data come from. Furthermore, we have proven that the
normalization factor in the variance estimator formula should be
population is known, and should be

if the true mean of the

if the mean itself also has to be estimated.

Feature extraction using PCA

Contents [hide] [hide]

1 Introduction
2 PCA as a decorrelation method
3 PCA as an orthogonal regression method
4 A practical PCA application: Eigenfaces
5 The PCA recipe
o 5.1 1) Center the data
o 5.2 2) Normalize the data
o 5.3 3) Calculate the eigendecomposition
o 5.4 4) Project the data
6 PCA pitfalls
7 Source Code
8 Conclusion

Introduction
In this article, we discuss how Principal Component Analysis (PCA) works, and how it can be used as a
dimensionality reduction technique for classification problems. At the end of this article, Matlab source
code is provided for demonstration purposes.
In an earlier article, we discussed the so called Curse of Dimensionality and showed that classifiers tend to
overfit the training data in high dimensional spaces. The question then rises which features should be
preferred and which ones should be removed from a high dimensional feature vector.
If all features in this feature vector were statistically independent, one could simply eliminate the least
discriminative features from this vector. The least discriminative features can be found by various
greedy feature selection approaches. However, in practice, many features depend on each other or on an
underlying unknown variable. A single feature could therefore represent a combination of multiple types of
information by a single value. Removing such a feature would remove more information than needed. In
the next paragraphs, we introduce PCA as a feature extraction solution to this problem, and introduce its
inner workings from two different perspectives.

PCA as a decorrelation method

More often than not, features are correlated. As an example, consider the case where we want to use the
red, green and blue components of each pixel in an image to classify the image (e.g. detect dogs versus
cats). Image sensors that are most sensitive to red light also capture some blue and green light. Similarly,
sensors that are most sensitive to blue and green light also exhibit a certain degree of sensitivity to red

light. As a result, the R, G, B components of a pixel are statistically correlated. Therefore, simply eliminating
the R component from the feature vector, also implicitly removes information about the G and B channels.
In other words, before eliminating features, we would like to transform the complete feature space such
that the underlying uncorrelated components are obtained.
Consider the following example of a 2D feature space:

Figure 1 2D Correlated data with eigenvectors shown in color.

The features

and , illustrated by figure 1, are clearly correlated. In fact, their covariance matrix is:

In an earlier article we discussed the geometric interpretation of the covariance matrix. We saw that the
covariance matrix can be decomposed as a sequence of rotation and scaling operations on white,
uncorrelated data, where the rotation matrix is defined by theeigenvectors of this covariance matrix.
Therefore, intuitively, it is easy to see that the data shown in figure 1 can be decorrelated by rotating
each data point such that the eigenvectors become the new reference axes:

(1)

Figure 2.2D Uncorrelated data with eigenvectors shown in color.

The covariance matrix of the resulting data is now diagonal, meaning that the new axes are uncorrelated:

In fact, the original data used in this example and shown by figure 1 was generated by linearly combining
two 1D Gaussian feature vectors
and
as follows:

Since the features and are linear combinations of some unknown underlying
components and , directly eliminating either or as a feature would have removed some
information from both and . Instead, rotating the data by the eigenvectors of its covariance matrix,
allowed us to directly recover the independent components and (up to a scaling factor). This can
be seen as follows: The eigenvectors of the covariance matrix of the original data are (each column
represents an eigenvector):

The first thing to notice is that in this case is a rotation matrix, corresponding to a rotation of 45 degrees
(cos(45)=0.7071), which indeed is evident from figure 1. Secondly, treating as a linear transformation
matrix results in a new coordinate system, such that each new feature and is expressed as a linear
combination of the original features and :

(2)
and

(3)
In other words, decorrelation of the feature space corresponds to the recovery of the unknown,
uncorrelated components and of the data (up to an unknown scaling factor if the transformation
matrix was not orthogonal). Once these components have been recovered, it is easy to reduce the
dimensionality of the feature space by simply eliminating either or .
In the above example we started with a two-dimensional problem. If we would like to reduce the
dimensionality, the question remains whether to eliminate (and thus ) or (and thus ).
Although this choice could depend on many factors such as the separability of the data in case of
classification problems, PCA simply assumes that the most interesting feature is the one with the largest

variance or spread. This assumption is based on an information theoretic point of view, since the dimension
with the largest variance corresponds to the dimension with the largest entropy and thus encodes the
most information. The smallest eigenvectors will often simply represent noise components, whereas the
largest eigenvectors often correspond to the principal components that define the data.
Dimensionality reduction by means of PCA is then accomplished simply by projecting the data onto the
largest eigenvectors of its covariance matrix. For the above example, the resulting 1D feature space is
illustrated by figure 3:

Figure 3. PCA: 2D data projected onto its largest eigenvector.

Obivously, the above example easily generalizes to higher dimensional feature spaces. For instance, in the
three-dimensional case, we can either project the data onto the plane defined by the two largest
eigenvectors to obtain a 2D feature space, or we can project it onto the largest eigenvector to obtain a 1D
feature space. This is illustrated by figure 4:

Figure 4. 3D data projected onto a 2D or 1D linear subspace by means of Principal Component Analysis.
In general, PCA allows us to obtain a linear M-dimensional subspace of the original N-dimensional data,
where
. Furthermore, if the unknown, uncorrelated components are Gaussian distributed, then
PCA actually acts as an independent component analysis since uncorrelated Gaussian variables are
statistically independent. However, if the underlying components are not normally distributed, PCA merely
generates decorrelated variables which are not necessarily statistically independent. In this case, non-linear
dimensionality reduction algorithms might be a better choice.

PCA as an orthogonal regression method

In the above discussion, we started with the goal of obtaining independent components (or at least
uncorrelated components if the data is not normally distributed) to reduce the dimensionality of the
feature space. We found that these so called principal components are obtained by the
eigendecomposition of the covariance matrix of our data. The dimensionality is then reduced by projecting
the data onto the largest eigenvectors.
Now lets forget about our wish to find uncorrelated components for a while. Instead, we will now try to
reduce the dimensionality by finding a linear subspace of the original feature space onto which we can
project our data such that the projection error is minimized. In the 2D case, this means that we try to find a

vector such that projecting the data onto this vector corresponds to a projection error that is lower than
the projection error that would be obtained when projecting the data onto any other possible vector. The
question is then how to find this optimal vector.
Consider the example shown by figure 5. Three different projection vectors are shown, together with the
resulting 1D data. In the next paragraphs, we will discuss how to determine which projection vector
minimizes the projection error. Before searching for a vector that minimizes the projection error, we have to
define this error function.

Figure 5 Dimensionality reduction by projection onto a linear subspace

A well known method to fit a line to 2D data is least squares regression. Given the independent
variable and the dependent variable , the least squares regressor corresponds to the
line

, such that the sum of the squared residual errors

is

minimized. In other words, if is treated as the independent variable, then the obtained
regressor
is a linear function that can predict the dependent variable such that the squared error
is minimal. The resulting model
is illustrated by the blue line in figure 5, and the error that is
minimized is illustrated in figure 6.

Figure 6. Linear regression where x is the independent variable and y is the dependent variable,
corresponds to minimizing the vertical projection error.
However, in the context of feature extraction, one might wonder why we would define feature as the
independent variable and feature as the dependent variable. In fact, we could easily define as the
independent variable and find a linear function
that predicts the dependent variable , such
that
is minimized. This corresponds to minimization of the horizontal projection
error and results in a different linear model as shown by figure 7:

Figure 7. Linear regression where y is the independent variable and x is the dependent variable,
corresponds to minimizing the horizontal projection error.
Clearly, the choice of independent and dependent variables changes the resulting model, making ordinary
least squares regression an asymmetric regressor. The reason for this is that least squares regression
assumes the independent variable to be noise-free, whereas the dependent variable is assumed to be
noisy. However, in the case of classification, all features are usually noisy observations such that
neither or should be treated as independent. In fact, we would like to obtain a model
that
minimizes both the horizontal and the vertical projection error simultaneously. This corresponds to finding
a model such that the orthogonal projection error is minimized as shown by figure 8.

Figure 8. Linear regression where both variables are independent corresponds to minimizing the orthogonal
projection error.
The resulting regression is called Total Least Squares regression or orthogonal regression, and assumes that
both variables are imperfect observations. An interesting observation is now that the obtained vector,
representing the projection direction that minimizes the orthogonal projection error, corresponds the the
largest principal component of the data:

Figure 9. The vector which the data can be projected unto with minimal orthogonal error corresponds to
the largest eigenvector of the covariance matrix of the data.
In other words, if we want to reduce the dimensionality by projecting the original data onto a vector such
that the squared projection error is minimized in all directions, we can simply project the data onto the
largest eigenvectors. This is exactly what we called Principal Component Analysis in the previous section,
where we showed that such projection also decorrelates the feature space.

A practical PCA application: Eigenfaces

Although the above examples are limited to two or three dimensions for visualization purposes,
dimensionality reduction usually becomes important when the number of features is not negligible
compared to the number of training samples. As an example, suppose we would like to perform face
recognition, i.e. determine the identity of the person depicted in an image, based on a training dataset of
labeled face images. One approach might be to treat the brightness of each pixel of the image as a feature.
If the input images are of size 3232 pixels, this means that the feature vector contains 1024 feature values.
Classifying a new face image can then be done by calculating the Euclidean distance between this 1024-

dimensional vector, and the feature vectors of the people in our training dataset. The smallest distance
then tells us which person we are looking at.
However, operating in a 1024-dimensional space becomes problematic if we only have a few hundred
training samples. Furthermore, Euclidean distances behave strangely in high dimensional spaces as
discussed in an earlier article. Therefore, we could use PCA to reduce the dimensionality of the feature
space by calculating the eigenvectors of the covariance matrix of the set of 1024-dimensional feature
vectors, and then projecting each feature vector onto the largest eigenvectors.
Since the eigenvector of 2D data is 2-dimensional, and an eigenvector of 3D data is 3-dimensional, the
eigenvectors of 1024-dimensional data is 1024-dimensional. In other words, we could reshape each of the
1024-dimensional eigenvectors to a 3232 image for visualization purposes. Figure 10 shows the first four
eigenvectors obtained by eigendecomposition of the Cambridge face dataset:

Figure 10. The four largest eigenvectors, reshaped to images, resulting in so called EigenFaces.
(source:https://nl.wikipedia.org/wiki/Eigenface)
Each 1024-dimensional feature vector (and thus each face) can now be projected onto the N largest
eigenvectors, and can be represented as a linear combination of these eigenfaces. The weights of these
linear combinations determine the identity of the person. Since the largest eigenvectors represent the
largest variance in the data, these eigenfaces describe the most informative image regions (eyes, noise,
mouth, etc.). By only considering the first N (e.g. N=70) eigenvectors, the dimensionality of the feature
space is greatly reduced.

The remaining question is now how many eigenfaces should be used, or in the general case; how many
eigenvectors should be kept. Removing too many eigenvectors might remove important information from the
feature space, whereas eliminating too few eigenvectors leaves us with the curse of dimensionality.
Regrettably there is no straight answer to this problem. Although cross-validation techniques can be used
to obtain an estimate of this hyperparameter, choosing the optimal number of dimensions remains a
problem that is mostly solved in an empirical (an academic term that means not much more than trialand-error) manner. Note that it is often useful to check how much (as a percentage) of the variance of the
original data is kept while eliminating eigenvectors. This is done by dividing the sum of the kept eigenvalues
by the sum of all eigenvalues.

The PCA recipe

Based on the previous sections, we can now list the simple recipe used to apply PCA for feature extraction:

1) Center the data

In an earlier article, we showed that the covariance matrix can be written as a sequence of linear
operations (scaling and rotations). The eigendecomposition extracts these transformation matrices: the
eigenvectors represent the rotation matrix, while the eigenvalues represent the scaling factors. However,
the covariance matrix does not contain any information related to the translation of the data. Indeed, to
represent translation, an affine transformation would be needed instead of a linear transformation.
Therefore, before applying PCA to rotate the data in order to obtain uncorrelated axes, any existing shift
needs to be countered by subtracting the mean of the data from each data point. This simply corresponds
to centering the data such that its average becomes zero.

2) Normalize the data

The eigenvectors of the covariance matrix point in the direction of the largest variance of the data.
However, variance is an absolute number, not a relative one. This means that the variance of data,
measured in centimeters (or inches) will be much larger than the variance of the same data when
measured in meters (or feet). Consider the example where one feature represents the length of an object
in meters, while the second feature represents the width of the object in centimeters. The largest variance,
and thus the largest eigenvector, will implicitly be defined by the first feature if the data is not normalized.
To avoid this scale-dependent nature of PCA, it is useful to normalize the data by dividing each feature by
its standard deviation. This is especially important if different features correspond to different metrics.

3) Calculate the eigendecomposition

Since the data will be projected onto the largest eigenvectors to reduce the dimensionality,
the eigendecomposition needs to be obtained. One of the most widely used methods to efficiently
calculate the eigendecomposition is Singular Value Decomposition (SVD).

4) Project the data

To reduce the dimensionality, the data is simply projected onto the largest eigenvectors. Let be the
matrix whose columns contain the largest eigenvectors and let be the original data whose columns
contain the different observations. Then the projected data
is obtained as
. We can
either choose the number of remaining dimensions, i.e. the columns of , directly, or we can define the
amount of variance of the original data that needs to kept while eliminating eigenvectors. If
only eigenvectors are kept, and
represent the corresponding eigenvalues, then the amount
of variance that remains after projecting the original -dimensional data can be calculated as:

(4)

PCA pitfalls
In the above discussion, several assumptions have been made. In the first section, we discussed how PCA
decorrelates the data. In fact, we started the discussion by expressing our desire to recover the unknown,
underlying independent components of the observed features. We then assumed that our data was
normally distributed, such that statistical independence simply corresponds to the lack of a linear
correlation. Indeed, PCA allows us to decorrelate the data, thereby recovering the independent
components in case of Gaussianity. However, it is important to note that decorrelation only corresponds to
statistical independency in the Gaussian case. Consider the data obtained by sampling half a period
of
:

Figure 11 Uncorrelated data is only statistically independent if normally distributed. In this example a clear
non-linear dependency still exists: y=sin(x).
Although the above data is clearly uncorrelated (on average, the y-value increases as much as it decreases
when the x-value goes up) and therefore corresponds to a diagonal covariance matrix, there still is a clear
non-linear dependency between both variables.
In general, PCA only uncorrelates the data but does not remove statistical dependencies. If the underlying
components are known to be non-Gaussian, techniques such as ICA could be more interesting. On the
other hand, if non-linearities clearly exist, dimensionality reduction techniques such as non-linear PCA can
be used. However, keep in mind that these methods are prone to overfitting themselves, since more
parameters are to be estimated based on the same amount of training data.
A second assumption that was made in this article, is that the most discriminative information is captured
by the largest variance in the feature space. Since the direction of the largest variance encodes the most
information this is likely to be true. However, there are cases where the discriminative information actually
resides in the directions of the smallest variance, such that PCA could greatly hurt classification
performance. As an example, consider the two cases of figure 12, where we reduce the 2D feature space to
a 1D representation:

Figure 12. In the first case, PCA would hurt classification performance because the data becomes linearly
unseparable. This happens when the most discriminative information resides in the smaller eigenvectors.
If the most discriminative information is contained in the smaller eigenvectors, applying PCA might actually
worsen the Curse of Dimensionality because now a more complicated classification model (e.g. non-linear
classifier) is needed to classify the lower dimensional problem. In this case, other dimensionality reduction
methods might be of interest, such asLinear Discriminant Analysis (LDA) which tries to find the projection
vector that optimally separates the two classes.

Source Code
The following code snippet shows how to perform principal component analysis for dimensionality
reduction in Matlab:
Matlab source code

Conclusion
In this article, we discussed the advantages of PCA for feature extraction and dimensionality reduction from
two different points of view. The first point of view explained how PCA allows us to decorrelate the feature
space, whereas the second point of view showed that PCA actually corresponds to orthogonal regression.
Furthermore, we briefly introduced Eigenfaces as a well known example of PCA based feature extraction,
and we covered some of the most important disadvantages of Principal Component Analysis.

What are eigenvectors and eigenvalues?

Contents [hide] [hide]

1 Introduction
2 Calculating the eigenvalues
3 Calculating the first eigenvector
4 Calculating the second eigenvector
5 Conclusion

Introduction
Eigenvectors and eigenvalues have many important applications in computer vision and machine learning
in general. Well known examples are PCA (Principal Component Analysis)for dimensionality reduction
or EigenFaces for face recognition. An interesting use of eigenvectors and eigenvalues is also illustrated in
my post about error ellipses. Furthermore, eigendecomposition forms the base of the geometric
interpretation of covariance matrices, discussed in an more recent post. In this article, I will provide a
gentle introduction into this mathematical concept, and will show how to manually obtain the
eigendecomposition of a 2D square matrix.
An eigenvector is a vector whose direction remains unchanged when a linear transformation is applied to it.
Consider the image below in which three vectors are shown. The green square is only drawn to illustrate
the linear transformation that is applied to each of these three vectors.

Eigenvectors (red) do not change direction when a linear transformation (e.g. scaling) is applied to them.
Other vectors (yellow) do.
The transformation in this case is a simple scaling with factor 2 in the horizontal direction and factor 0.5 in
the vertical direction, such that the transformation matrix is defined as:

.
A vector
is then scaled by applying this transformation as
. The above figure
shows that the direction of some vectors (shown in red) is not affected by this linear transformation. These
vectors are called eigenvectors of the transformation, and uniquely define the square matrix . This
unique, deterministic relation is exactly the reason that those vectors are called eigenvectors (Eigen
means specific in German).
In general, the eigenvector

of a matrix

is the vector for which the following holds:

(1)
where is a scalar value called the eigenvalue. This means that the linear transformation
vector is completely defined by .

on

We can rewrite equation (1) as follows:

(2)
where is the identity matrix of the same dimensions as

However, assuming that is not the null-vector, equation (2) can only be defined if
is not
invertible. If a square matrix is not invertible, that means that itsdeterminant must equal zero. Therefore, to
find the eigenvectors of , we simply have to solve the following equation:

(3)
In the following sections we will determine the eigenvectors and eigenvalues of a matrix
equation (3). Matrix in this example, is defined by:

(4)

, by solving

Calculating the eigenvalues

To determine the eigenvalues for this example, we substitute
obtain:

in equation (3) by equation (4) and

(5)
Calculating the determinant gives:

(6)
To solve this quadratic equation in , we find the discriminant:

Since the discriminant is strictly positive, this means that two different values for

(7)

exist:

We have now determined the two eigenvalues and . Note that a square matrix of
size
always has exactly eigenvalues, each with a corresponding eigenvector. The eigenvalue
specifies the size of the eigenvector.

Calculating the first eigenvector

We can now determine the eigenvectors by plugging the eigenvalues from equation (7) into equation (1)
that originally defined the problem. The eigenvectors are then found by solving this system of equations.
We first do this for eigenvalue

, in order to find the corresponding first eigenvector:

Since this is simply the matrix notation for a system of equations, we can write it in its equivalent form:

(8)
and solve the first equation as a function of

, resulting in:

(9)
Since an eigenvector simply represents an orientation (the corresponding eigenvalue represents the
magnitude), all scalar multiples of the eigenvector are vectors that are parallel to this eigenvector, and are
therefore equivalent (If we would normalize the vectors, they would all be equal). Thus, instead of further
solving the above system of equations, we can freely chose a real value for either
or
, and
determine the other one by using equation (9).
For this example, we arbitrarily choose
corresponds to eigenvalue
is

, such that

. Therefore, the eigenvector that

(10)

Calculating the second eigenvector

Calculations for the second eigenvector are similar to those needed for the first eigenvector;
We now substitute eigenvalue
into equation (1), yielding:

(11)
Written as a system of equations, this is equivalent to:

(12)
Solving the first equation as a function of

resuls in:

(13)
We then arbitrarily choose
eigenvalue
is

(14)

, and find

. Therefore, the eigenvector that corresponds to

Conclusion
In this article we reviewed the theoretical concepts of eigenvectors and eigenvalues. These concepts are of
great importance in many techniques used in computer vision and machine learning, such as
dimensionality reduction by means of PCA, or face recognition by means of EigenFaces.