Chapter 4

Variances and covariances

4.1 Overview
The expected value of a random variable gives a crude measure for the
“center of location” of the distribution of that random variable. For instance,
if the distribution is symmetric about a value µ then the expected value
equals µ. To refine the picture of a distribution about its “center of location”
we need some measure of spread (or concentration) around that value. For
many distributions the simplest measure to calculate is the variance (or,
more precisely, the square root of the variance).

Definition. The variance of a random variable
X with expected value

2
EX = µ is defined as var(X) = E (X − µ) . The square root of the
variance of a random variable is called its standard deviation, sometimes
denoted by sd(X).

The variance of a random variable X is unchanged by an added constant:

var(X + C) = var(X) for every constant C, because (X + C) − E(X + C) =
X − EX, the C’s cancelling. It is a desirable property that the spread
should not be affected by a change in location. However, it is also desirable
that multiplication by a constant should change the spread: var(CX) =
C 2 var(X) and sd(CX) = |C|sd(X), because (CX − E(CX))2 = C 2 (X −
EX)2 . In summary: for constants a and b,

var(a + bX) = b2 var(X) and sd(a + bX) = |b|sd(X).

Statistics 241/541 fall 2014 David

c Pollard, Sept2014 1
4. Variances and covariances 2

Remark. Try not to confuse properties of expected values with

properties of variances: for constants a and b we have var(a + bX) =
b2 var(X) but E(a + bX) = a + bEX. Measures of location (expected
value) and spread (standard deviation) should react differently to linear
transformations of the variable. As another example: if a given piece of
“information” implies that a random variable X must take the constant
value C then E(X | information) = C, but var(X | information) = 0.
It is a common blunder to confuse the formula for the variance of
a difference with the formula E(Y − Z) = EY − EZ. If you ever find
yourself wanting to assert that var(Y − Z) is equal to var(Y ) − var(Z),
think again. What would happen if var(Z) were larger than var(Y )?
Variances can’t be negative.

There is an enormous probability literature that deals with approxima-

tions to distributions, and bounds for probabilities, expressible in terms of
expected values and variances. One of the oldest and simplest examples,
the Tchebychev inequality, is still useful, even though it is rather crude by
modern standards.

Example <4.1> The Tchebychev inequality: P{|X−µ| ≥ } ≤ var(X)/2 ,

where µ = EX and  > 0.

Remark. In the Chapter on the normal distribution you will find more
refined probability approximations involving the variance.

The Tchebychev inequality gives the right insight when dealing with
sums of random variables, for which variances are easy to calculate. Sup-
pose EY = µY and EZ = µZ . Then

var(Y + Z) = E [Y − µY + Z − µZ ]2
= E (Y − µY )2 + 2(Y − µY )(Z − µZ ) + (Z − µZ )2
 

= var(Y ) + 2cov(Y, Z) + var(Z)

where cov(Y, Z) denotes the covariance between Y and Z:

cov(Y, Z) := E [(Y − µY )(Z − µZ )] .

Remark. Notice that cov(X, X) = var(X). Results about covariances

contain results about variances as special cases.

Statistics 241/541 fall 2014 David

c Pollard, Sept2014
 4. Variances and covariances 3

More generally, for constants a, b, c, d, and random variables U, V, Y, Z,

cov(aU + bV, cY + dZ)

= ac cov(U, Y ) + bc cov(V, Y ) + ad cov(U, Z) + bd cov(V, Z).

It is easier to see the pattern if we work with the centered random variables
U 0 = U − µU , . . . , Z 0 = Z − µZ . For then the left-hand side becomes

E (aU 0 + bV 0 )(cY 0 + dZ 0 )
 

= E ac U 0 Y 0 + bc V 0 Y 0 + ad U 0 Z 0 + bd V 0 Z 0
 

= ac E(U 0 Y 0 ) + bc E(V 0 Y 0 ) + ad E(U 0 Z 0 ) + bd E(V 0 Z 0 ).

The expected values in the last line correspond to the four covariances.
Sometimes it is easier to subtract off the expected values at the end of
the calculation, by means of the formulae cov(Y, Z) = E(Y Z) − (EY )(EZ)
and, as a particular case, var(X) = E(X 2 ) − (EX)2 . Both formulae follow
via an expansion of the product:

cov(Y, Z) = E (Y Z − µY Z − µZ Y + µY µZ )
= E(Y Z) − µY EZ − µZ EY + µY µZ
= E(Y Z) − µY µZ .

Rescaled covariances define correlations, a concept that is much abused

by those who do not understand probability.

Definition. The correlation between Y and Z is defined as

cov(Y, Z)
corr(Y, Z) = q
var(Y )var(Z)

The random variables Y and Z are said to be uncorrelated if corr(Y, Z) = 0.

Remark. Strictly speaking, the variance of a random variable is not

well defined unless it has a finite expectation. Similarly, we should not
talk about corr(Y, Z) unless both random variables have well defined
variances for which 0 < var(Y ) < ∞ and 0 < var(Z) < ∞.

Example <4.2> When well defined, correlations always lie between +1

and −1.

Statistics 241/541 fall 2014 David

c Pollard, Sept2014
4. Variances and covariances 4

Variances for sums of uncorrelated random variables grow more slowly

than might be anticipated. If Y and Z are uncorrelated, the covariance
term drops out from the expression for the variance of their sum, leaving
var(Y +Z) = var(Y )+var(Z). Similarly, if X1 , . . . , Xn are random variables
for which cov(Xi , Xj ) = 0 for each i 6= j then

var(X1 + · · · + Xn ) = var(X1 ) + · · · + var(Xn )

You should check the last assertion by expanding out the quadratic in the
variables Xi − EXi , observing how all the cross-product terms disappear
because of the zero covariances. These facts lead to a useful concentration
property.

Example <4.3> Concentration of averages around expected value

Zero correlation is often deduced from independence. A pair of random

variables X and Y is said to be independent if every event determined by X
is independent of every event determined by Y . For example, independence
implies that events such as {X ≤ 5} and {7 ≤ Y ≤ 18} are independent,
and so on. Independence of the random variables also implies independence
of functions of those random variables. For example, sin(X) would be inde-
pendent of eY , and so on. For the purposes of Stat241, you should not fret
about the definition of independence: Just remember to explain why you re-
gard some pieces of information as irrelevant when you calculate conditional
probabilities and conditional expectations.
For example, suppose a random variable X can take values x1 , x2 , . . .
and that X is independent of another random variable Y . Consider the
expected value of a product g(X)h(Y ), for any functions g and h. Calculate
by conditioning on the possible values taken by X:
X
Eg(X)h(Y ) = P{X = xi }E(g(X)h(Y ) | X = xi ).
i

Given that X = xi , we know that g(X) = g(xi ) but we get no help with
understanding the behavior of h(Y ). Thus, independence implies

E(g(X)h(Y ) | X = xi ) = g(xi )E(h(Y ) | X = xi ) = g(xi )Eh(Y ).

Deduce that
X
Eg(X)h(Y ) = P{X = xi }g(xi )Eh(Y ) = Eg(X)Eh(Y ).
i

Statistics 241/541 fall 2014 David

c Pollard, Sept2014
 4. Variances and covariances 5

Put another way, if X and Y are independent random variables

cov g(X), h(Y ) = E g(X)h(Y ) − (Eg(X)) (Eh(Y )) = 0.

That is, each function of X is uncorrelated with each function of Y . In

particular, if X and Y are independent then they are uncorrelated. The
converse is not usually true: uncorrelated random variables need not be
independent.

Example <4.4> An example of uncorrelated random variables that are

dependent

The concentration phenomenon can also hold for averages of dependent

random variables.

Example <4.5> Comparison of spread in sample averages for sampling

with and without replacement: the Decennial Census.

As with expectations, variances and covariances can also be calculated

conditionally on various pieces of information. The conditioning formula in
the final Example has the interpretation of a decomposition of “variability”
into distinct sources, a precursor to the statistical technique know as the
“analysis of variance”.

Example <4.6> An example to show how variances can sometimes be

decomposed into components attributable to difference sources. (Can be
skipped.)

4.2 Things to remember

• Eg(X)h(Y ) = Eg(X)Eh(Y ) if X and Y are independent random vari-
ables

• the definitions of variance and covariance, and their expanded forms

cov(Y, Z) = E(Y Z) − (EY )(EZ) and var(X) = E(X 2 ) − (EX)2

• var(a + bX) = b2 var(X) and sd(a + bX) = |b|sd(X) for constants a

and b.

Statistics 241/541 fall 2014 David

c Pollard, Sept2014
 4. Variances and covariances 6

• For constants a, b, c, d, and random variables U, V, Y, Z,

cov(aU + bV, cY + dZ)
= ac cov(U, Y ) + bc cov(V, Y ) + ad cov(U, Z) + bd cov(V, Z).

• Sampling without replacement gives smaller variances than sampling

with replacement.

4.3 The examples

<4.1> Example. The Tchebychev inequality asserts: for a random variable X
with expected value µ,
P{|X − µ| > } ≤ var(X)/2 for each  > 0.
The inequality becomes obvious if we write F for the event {|X − µ| > }.
First note that IF ≤ |X −µ|2 /2 : when IF = 0 the inequality holds for trivial
reasons; and when IF takes the value one, the random variable |X −µ|2 must
be larger than 2 . It follows that
P{|X − µ| > } = PF = EIF ≤ E|X − µ|2 /2 .

<4.2> Example. When well defined, correlations always lies between +1 and −1.
Suppose
EY = µY and var(Y ) = σY2
EZ = µY and var(Z) = σZ2
Define standardized variables
Y − µY Z − µZ
Y0 = and Z0 = .
σY σZ
Note that EY 0 = EZ 0 = 0 and var(Y 0 ) = var(Z 0 ) = 1. Also
corr(Y, Z) = cov(Y 0 Z 0 ) = E(Y 0 Z 0 ).
Use the fact that variances are always nonnegative to deduce that
0 ≤ var(Y 0 + Z 0 ) = var(Y 0 ) + 2cov(Y 0 , Z 0 ) + var(Z 0 ) = 2 + 2cov(Y 0 , Z 0 ),
which rearranges to cov(Y 0 , Z 0 ) ≥ −1. Similarly
0 ≤ var(Y 0 − Z 0 ) = var(Y 0 ) − 2cov(Y 0 , Z 0 ) + var(Z 0 ) = 2 − 2cov(Y 0 , Z 0 ),
which rearranges to cov(Y 0 , Z 0 ) ≤ +1. 

Statistics 241/541 fall 2014 David

c Pollard, Sept2014
 4. Variances and covariances 7

<4.3> Example. Suppose X1 , . . . , Xn are uncorrelated random variables, each

with expected value µ and variance σ 2 . By repeated application of the
formula for the variance of a sum of variables with zero covariances,
var (X1 + · · · + Xn ) = var(X1 ) + · · · + var(Xn ) = nσ 2 .
Typically the Xi would come from repeated independent measurements of
some unknown quantity. The random variable X = (X1 + · · · + Xn )/n is
then called the sample mean.
The variance of the sample mean decreases like 1/n,
var(X) = (1/n)2 var (X1 + · · · + Xn ) = σ 2 /n.
From the Tchebychev inequality,
P{|X − µ| > } ≤ (σ 2 /n)/2 for each  > 0.
In particular, for each positive constant C,
√
P{|X − µ| > Cσ/ n} ≤ 1/C 2 .
√
For example, there is at most a 1% chance that X lies more than 10σ/ n
away from µ. (A normal approximation will give a much tighter bound.)
Note well the dependence on n. 

<4.4> Example. Consider two independent rolls of a fair die. Let X denote the
value rolled the first time and Y denote the value rolled the second time.
The random variables X and Y are independent, and they have the same
distribution. Consequently cov(X, Y ) = 0, and var(X) = var(Y ).
The two random variables X + Y and X − Y are uncorrelated:
cov(X + Y, X − Y )
= cov(X, X) + cov(X, −Y ) + cov(Y, X) + cov(Y, −Y )
= var(X) − cov(X, Y ) + cov(Y, X) − var(Y )
= 0.
Nevertheless, the sum and difference are not independent. For example,
1
P{X + Y = 12} = P{X = 6}P{Y = 6} =
36
but
P{X + Y = 12 | X − Y = 5} = P{X + Y = 12 | X = 6, Y = 1} = 0.


Statistics 241/541 fall 2014 David

c Pollard, Sept2014
 4. Variances and covariances 8

<4.5> Example. Until quite recently, in the Decennial Census of Housing and
Population the Census Bureau would obtain some more detailed about the
population via information from a more extensive list of questions sent to
only a random sample of housing units. For an area like New Haven, about
1 in 6 units would receive the so-called “long form”.
For example, one question on the long form asked for the number of
rooms in the housing unit. We could imagine the population of all units
numbered 1, 2, . . . , N , with the ith unit containing yi rooms. Complete
enumeration would reveal the value of the population average,
1
ȳ = (y1 + y2 + · · · + yN ) .
N
A sample can provide a good estimate of ȳ with less work.
Suppose a sample of n housing units is selected from the population
without replacement. (For the Decennial Census, n ≈ N/6.) The answer
from each unit is a random variable that could take each of the values
y1 , y2 , . . . , yN , each with probability 1/N .
Remark. It might be better to think of a random variable that takes
each of the values 1, 2, . . . , N with probability 1/N , then take the
corresponding number of rooms as the value of the random variable
that is recorded. Otherwise we can fall into verbal ambiguities when
many of the units have the same number of rooms.

That is, the sample consists of random variables Y1 , Y2 , . . . , Yn , for each of

which
1
P{Yi = yj } = for j = 1, 2, . . . , N .
N
Notice that
1 XN
EYi = yj = ȳ,
N j=1

and consequently, the sample average Ȳ = (Y1 +· · ·+Yn )/n also has expected
value ȳ. Notice also that each Yi has the same variance,
1 XN
var(Yi ) = (yj − ȳ)2 ,
N j=1

a quantity that I will denote by σ 2 .

If the sample is taken without replacement—which, of course, the Census
Bureau had to do, if only to avoid media ridicule—the random variables are

Statistics 241/541 fall 2014 David

c Pollard, Sept2014
 4. Variances and covariances 9

dependent. For example, in the extreme case where n = N , we would

necessarily have
Y1 + Y2 + · · · + YN = y1 + y2 + · · · + yN ,
so that YN would be a function of the other Yi ’s, a most extreme form of
dependence. Even if n < N , there is still some dependence, as you will soon
see.
Sampling with replacement would be mathematically simpler, for then
the random variables Y
i would be independent, and, as in Example <4.3>,
we would have var Ȳ = σ 2 /n. With replacement, it is possible that the
same unit might be sampled more than once, especially if the sample size is
an appreciable fraction of the population size. There is also some ineffici-
ciency in sampling with replacement, as shown by a calculation of variance
for sampling without replacement:


2
var Ȳ = E Ȳ − ȳ
 X 2
1 n
=E (Yi − ȳ)
n i=1
1  X n X 
= 2E (Yi − ȳ)2 + 2 (Yi − ȳ)(Yj − ȳ)
n i=1 1≤i<j≤n
1 X n X 
= 2 E (Yi − ȳ)2 + 2 E ((Yi − ȳ)(Yj − ȳ))
n i=1 1≤i<j≤n
1 X n X 
= 2 var(Yi ) + cov(Yi , Yj )
n i=1 1≤i6=j≤n

What formula did There are n variance terms and n(n − 1) covariance terms. We know that
I just rederive? each Yi has variance σ 2 , regardless of the dependence between the variables.
The effect of the dependence shows up in the covariance terms. By symme-
try, cov(Yi , Yj ) is the same for each pair i < j, a value that I will denote
by c. Thus, for sampling without replacement,
1
σ 2 (n − 1)c
var Ȳ = 2 nσ 2 + n(n − 1)c =


(∗) + .
n n n
We can calculate c directly, from the fact that the pair (Y1 , Y2 ) takes
each of N (N − 1) pairs of values (yi , yj ) with equal probability. Thus
1 X
c = cov(Y1 , Y2 ) = (yi − ȳ)(yj − ȳ).
N (N − 1) i6=j

If we added the “diagonal” terms (yi − ȳ)2 to the sum we would have the
expansion for the product
XN XN
(yi − ȳ) (yj − ȳ) ,
i=1 j=1

Statistics 241/541 fall 2014 David

c Pollard, Sept2014
 4. Variances and covariances 10

PN
which equals zero because N ȳ = i=1 yi . The expression for the covariance
simplifies to
σ2
 
1 2
XN
2
c = cov(Y1 , Y2 ) = 0 − (yi − ȳ) = − .
N (N − 1) i=1 N −1
Substitution in formula (∗) then gives
σ2 σ2 N − n
 
n−1
var(Ȳ ) = 1− = .
n N −1 n N −1
Compare with the σ 2 /n for var(Y ) under sampling with replacement.
The correction factor (N − n)/(N − 1) is close to 1 if the sample size n is
small compared with the population size N , but it can decrease the variance
of Y appreciably if n/N is not small. For example, if n ≈ N/6 (as with the
Census long form) the correction factor is approximately 5/6.
If n = N , the correction factor is zero. That is, var(Y ) = 0 if the
whole population is sampled. Indeed, when n = N we know that Ȳ equals
the population mean, ȳ, a constant. A random variable that always takes
the same constant value has zero variance. Thus the right-hand side of (∗)
must reduce to zero when we put n = N , which gives a quick method for
establishing the equality c = −σ 2 /(N − 1), without all the messing around
with sums of products and products of sums. 
<4.6> Example. Consider a two stage method for generating a random vari-
able Z. Suppose we have k different random variables Y1 , . . . , Yk , with
EYi = µi and var(Yi ) = σi2 . Suppose also that we have a random method
for selecting which variable to choose: a random variable X that is inde-
pendent of all the Yi ’s, with P{X = i} = pi for i = 1, 2, . . . , k, where
p1 + p2 + · · · + pk = 1. If X takes the value i, define Z to equal Yi .
The variability in Z is due to two effects: the variability of each Yi ; and
the variability of X. Conditional on X = i, we have Z equal to Yi , and
E (Z | X = i) = E(Yi ) = µi
var (Z | X = i) = E (Z − µi )2 | X = i = var(Yi ) = σi2 .

From the first formula we get

X X
EZ = P{X = i}E (Z | X = i) = pi µi ,
i i
a weighted average of the µi ’s that I will denote by µ̄. A similar conditioning
exercise gives
X  
var(Z) = E (Z − µ̄)2 = pi E (Z − µ̄)2 | X = i .
i

Statistics 241/541 fall 2014 David

c Pollard, Sept2014
4. Variances and covariances 11

If we could replace the µ̄ in the ith summand by µi , the sum would become a
weighted average of conditional variances. To achieve such an effect, rewrite
(Z − µ̄)2 as

(Z − µi + µi − µ̄)2 = (Z − µi )2 + 2(µi − µ̄)(Zi − µi ) + (µi − µ̄)2 .

Taking conditional expectations, we then get

 
E (Z − µ̄)2 | X = i
 
= E (Z − µ̄i )2 | X = i + 2(µi − µ̄)E (Z − µi | X = i) + (µi − µ̄)2 .

On the right-hand side, the first term equals σi2 , and the middle term disap-
pears because E(Z | X = i) = µi . With those simplifications, the expression
for the variance becomes
X X
var(Z) = pi σi2 + pi (µi − µ̄)2 .
i i

If we think of each Yi as coming from a separate “population”, the first

sum represents the component of variability within the populations, and the
second sum represents the variability between the populations.
The formula is sometimes written symbolically as

var(Z) = E (var(Z | X)) + var (E(Z | X)) ,

where E(Z | X) denotes the random variable that takes the value µi when X
takes the value i, and var(Z | X) denotes the random variable that takes
the value σi2 when X takes the value i. 

Statistics 241/541 fall 2014 David

c Pollard, Sept2014