18-753: Information Theory and Coding

Lecture 1: Introduction and Probability Review

Prof. Pulkit Grover

CMU

January 16, 2018

Bits and Information Theory

• More often than not, today’s information is measured in bits.

Bits and Information Theory

• More often than not, today’s information is measured in bits.

• Why?
Bits and Information Theory

• More often than not, today’s information is measured in bits.

• Why?
• Is it ok to represent signals as different as audio and video in the
same currency, bits?
Bits and Information Theory

• More often than not, today’s information is measured in bits.

• Why?
• Is it ok to represent signals as different as audio and video in the
same currency, bits?

• Also, why do I communicate over long distances with the exact same
currency?
 Video

• Video signals are made up of colors that

vary in space and time.

• Even if we are happy with pixels on a

screen how do we know that all these
colors are optimally described by bits?
 Audio

• Audio signals can be thought of

as an amplitude that varies
continuously with time.
15

• How can we optimally represent a

10

continuous signal in something 5

discrete like bits? 0

−5

−10
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Just a Single Bit

• These (and many other) signals can all be optimally represented by

bits.

• Information theory explains why and how.

• For now, let’s say we have a single bit that we really want to
communicate to our friend. Let’s say it represents whether or not it
is going to snow tonight.
Communication over Noisy Channels

• Let’s say we have to transmit our bit over the air using wireless
communication.

• We have a transmit antenna and our friend has a receive antenna.

• We’ll send out a negative amplitude (say −1) on our antenna when
the bit is 0 and a positive amplitude (say 1) when the bit is 1.

• Unfortunately, there are other signals in the air (natural and

man-made) so the receiver sees a noisy version of our transmission.

• If there are lots of these little effects, then the central limit theorem
tells us we can just model them as a single Gaussian random variable.
 Received Signal

• Here is the probability

1
distribution function for the
0.9

received signal. No longer a 0.8

clean −1 or 1. 0.7

0.6

• You can prove that the best 0.5

thing to do now is just decide 0.4

that it’s −1 if the signal is 0.3

below 0 and 1 otherwise. 0.2

0.1

• But that leaves us with a 0

−4 −3 −2 −1 0 1 2 3 4

probability of error.
 Received Signal

0.9

0.8

• One thing we can do is boost 0.7

our transmit power. 0.6

0.5

• The received signal will look 0.4

less and less noisy. 0.3

0.2

0.1

0
−4 −3 −2 −1 0 1 2 3 4
 Received Signal

0.9

0.8

• One thing we can do is boost 0.7

our transmit power. 0.6

0.5

• The received signal will look 0.4

less and less noisy. 0.3

0.2

0.1

0
−4 −3 −2 −1 0 1 2 3 4
Repetition Coding

• What if we can’t arbitrarily increase our transmit power?

• We can just repeat our bit many times! For example, if we have a 0,
just send −1, −1, −1, . . . , −1 and take a majority vote.

• Now we can get the probability of error to fall with the number of
repetitions.

• But the rate of incoming bits quickly goes to zero. Can we do

better?
Point-to-Point Communication

Zn

w Xn Yn ŵ
ENC DEC

• We know the capacity for an Gaussian channel:

1
C= log (1 + SNR) bits per channel use
2
• Proved by Claude Shannon in 1948.
• What does this mean?
 The Benefits of Blocklength

• We can’t predict what the noise is going

to be in a single channel use.

• But we do know that in the long run the

noise is going to behave a certain way.

• For example, if a given channel flips bits

with probability 0.1 then in the long run
1
approximately 10 of the bits will be
(Cover and Thomas,
flipped. Elements of Information Theory )
Capacity

• So if we are willing to allow for some delay we can communicate

reliably with positive rate!

• Capacity is actually very simple to calculate using the mutual

information, C = maxp(x) I(X; Y ).
What is Information Theory?

• A powerful mathematical framework that allows us to determine the

fundamental limits of information compression, storage, processing,
communication, and use.

• Provides the theoretical underpinnings as to why today’s networks

are completely digital.

• Unlike many other classes, we will strive to understand “why”

through full proofs.

• As initially formulated, information theory ignores semantics of the

message. We will touch upon how it is being extended to
incorporate semantics.
Why is it taught in an engineering department?

• Imagine if the proof that the speed of light c is the maximum

attainable speed also gave us a deep understanding of how to build a
starship that can approach this speed. And that we were able to
build such a starship within 50 years from the theory’s first
appearance!
Why is it taught in an engineering department?

• Imagine if the proof that the speed of light c is the maximum

attainable speed also gave us a deep understanding of how to build a
starship that can approach this speed. And that we were able to
build such a starship within 50 years from the theory’s first
appearance!

• That is exactly what information theory has given us for compression

and communication.
Why is it taught in an engineering department?

• Imagine if the proof that the speed of light c is the maximum

attainable speed also gave us a deep understanding of how to build a
starship that can approach this speed. And that we were able to
build such a starship within 50 years from the theory’s first
appearance!

• That is exactly what information theory has given us for compression

and communication.

• Example: Verizon spent 9.4 billion dollars on just 22MHz of

spectrum. Very important to them (and you) to get the highest
possible data rates.
Organizational Details

• This is 18-753: Information Theory and Coding.

• Taught by Prof. Pulkit Grover.
• Pre-requisites: Fluency in probability and mathematical maturity.
• Course Ingredients: 8 Homeworks, 2 Midterms, and a course project.
• Homework 1 out tonight, due next Thursday in class. (Just requires
probability.)
• Required textbook: Cover & Thomas, Elements of Information
Theory, 2nd edition.
Available online:
http://onlinelibrary.wiley.com/book/10.1002/047174882X (need to
login through CMU library)
• Office Hours: TBD.
Information Theory

• Consider the following (hypothetical) interactions between two

students in CMU.
• A: Have you ever been to the Museum of Natural History?
B: Yes.
Information Theory

• Consider the following (hypothetical) interactions between two

students in CMU.
• A: Have you ever been to the Museum of Natural History?
B: Yes.
• A: Have you ever been to the moon?
B: No.
Information Theory

• Consider the following (hypothetical) interactions between two

students in CMU.
• A: Have you ever been to the Museum of Natural History?
B: Yes.
• A: Have you ever been to the moon?
B: No.
• Both questions had two possible answers. Which interaction
conveyed more information?
Information Theory

• Consider the following (hypothetical) interactions between two

students in CMU.
• A: Have you ever been to the Museum of Natural History?
B: Yes.
• A: Have you ever been to the moon?
B: No.
• Both questions had two possible answers. Which interaction
conveyed more information?
• The “amount of information” in an event is related to how likely the
event is.
Compression and Communication

• Information Theory is the science of measuring of information.

Compression and Communication

• Information Theory is the science of measuring of information.

• This science has a profound impact on sensing, compression,
storage, extraction, processing, and communication of information.
• Compressing data such as audio, images, movies, text, sensor
measurements, etc. (Example: We will see the principle behind the
‘zip’ compression algorithm in this course.)
• Communicating data over noisy channels such as wires, wireless links,
memory (e.g. hard disks), etc.
Compression and Communication

• Information Theory is the science of measuring of information.

• This science has a profound impact on sensing, compression,
storage, extraction, processing, and communication of information.
• Compressing data such as audio, images, movies, text, sensor
measurements, etc. (Example: We will see the principle behind the
‘zip’ compression algorithm in this course.)
• Communicating data over noisy channels such as wires, wireless links,
memory (e.g. hard disks), etc.
• Specifically, we will be interested in determining the fundamental
limits of compression and communication. This will shed light on
how to engineer near-optimal systems.
Compression and Communication

• Information Theory is the science of measuring of information.

• This science has a profound impact on sensing, compression,
storage, extraction, processing, and communication of information.
• Compressing data such as audio, images, movies, text, sensor
measurements, etc. (Example: We will see the principle behind the
‘zip’ compression algorithm in this course.)
• Communicating data over noisy channels such as wires, wireless links,
memory (e.g. hard disks), etc.
• Specifically, we will be interested in determining the fundamental
limits of compression and communication. This will shed light on
how to engineer near-optimal systems.
• We will use probability as a “language” to describe and derive these
limits.
Compression and Communication

• Information Theory is the science of measuring of information.

• This science has a profound impact on sensing, compression,
storage, extraction, processing, and communication of information.
• Compressing data such as audio, images, movies, text, sensor
measurements, etc. (Example: We will see the principle behind the
‘zip’ compression algorithm in this course.)
• Communicating data over noisy channels such as wires, wireless links,
memory (e.g. hard disks), etc.
• Specifically, we will be interested in determining the fundamental
limits of compression and communication. This will shed light on
how to engineer near-optimal systems.
• We will use probability as a “language” to describe and derive these
limits.
• Information Theory has strong connections to Statistics, Physics,
Computer Science, and many other disciplines.
• This course is, to a degree, not about the meaning of information
(i.e. semantics). Although that is a philosophical undercurrent that
we will examine, question, understand.
General Setting

• Information Source: Data we want to send (e.g. a movie).

• Noisy Channel: Communication medium (e.g. a wire).
• Encoder: Maps source into a channel codeword (signal).
• Decoder: Reconstructs source from channel output (signal).
• Fidelity Criterion: Measures quality of the source reconstruction.

Noisy Source
Source Encoder Decoder
Channel Reconstruction
General Setting

• Information Source: Data we want to send (e.g. a movie).

• Noisy Channel: Communication medium (e.g. a wire).
• Encoder: Maps source into a channel codeword (signal).
• Decoder: Reconstructs source from channel output (signal).
• Fidelity Criterion: Measures quality of the source reconstruction.

Noisy Source
Source Encoder Decoder
Channel Reconstruction

• Goal: Transmit at the highest rate possible while meeting the

fidelity criterion.
• Example: Maximize frames/second while keeping mean-squared error
below 1%.
General Setting

• Information Source: Data we want to send (e.g. a movie).

• Noisy Channel: Communication medium (e.g. a wire).
• Encoder: Maps source into a channel codeword (signal).
• Decoder: Reconstructs source from channel output (signal).
• Fidelity Criterion: Measures quality of the source reconstruction.

Noisy Source
Source Encoder Decoder
Channel Reconstruction

• Goal: Transmit at the highest rate possible while meeting the

fidelity criterion.
• Example: Maximize frames/second while keeping mean-squared error
below 1%.
• Look Ahead: We will see a theoretical justification for the layered
protocol architecture of communication networks (combine optimal
compression with optimal communication)
Bits: The currency of information

• As we will see, bits are a universal currency of information (for single

sender, single receiver).
• Specifically, when we talk about sources, we often describe their size
in bits. Example: A small JPEG is around 100kB.
• Also, when we talk about channels, we often mention what rate they
can support. Example: A dial-up modem can send 14.4kB/sec.
High Dimensions

• Much of information theory relies on high-dimensional thinking.

That is, to compress and communicate data close to the
fundamental limits, we will need to operate over long blocklengths.
• This is on its face an extremely complex problem: nearly impossible
to “guess and check” solutions.
• Using probability, we will be able to reason about the existence (or
non-existence) of good schemes. This will give us insight into how
actually construct near-optimal schemes.
• Along the way, you will develop a lot of intuition for how
high-dimensional random vectors behave.
• We will now review some basic elements of probability that we will
need for the course.
Probability Review: Events

Elements of a Probability Space (Ω, F, P):

1 Sample space Ω = {ω1 , ω2 , . . .} whose elements correspond to the
possible outcomes.
2 Family of events F = {E1 , E2 , . . .} which is a collection of subsets
of Ω. We say that the event Ei occurs if the outcome ωj is an
element of Ei .
3 Probability measure P : F → R+ , a function that satisfies
(i) P(∅) = 0.
(ii) P(Ω) = 1.
(iii) If Ei ∩ Ej = ∅ (i.e. Ei and Ej are disjoint) for all i ̸= j, then
(∪
∞ ) ∞
∑
P Ei = P(Ei ) .
i=1 i=1
Probability Review: Events

Union of Events:
• P(E1 ∪ E2 ) = P(E1 ) + P(E2 ) − P(E1 ∩ E2 ). (Venn Diagram)
Probability Review: Events

Union of Events:
• P(E1 ∪ E2 ) = P(E1 ) + P(E2 ) − P(E1 ∩ E2 ). (Venn Diagram)
• More generally, we have the inclusion-exclusion principle:
(∪
n ) ∑
n ∑ ∑
P Ei = P(Ei ) − P(Ei ∩ Ej ) + P(Ei ∩ Ej ∩ Ek )
i=1 i=1 i<j i<j<k

∑ ( ∩
ℓ )
· · · + (−1) ℓ+1
P Eim + · · ·
i1 <i2 <···<iℓ m=1
Probability Review: Events

Union of Events:
• P(E1 ∪ E2 ) = P(E1 ) + P(E2 ) − P(E1 ∩ E2 ). (Venn Diagram)
• More generally, we have the inclusion-exclusion principle:
(∪
n ) ∑
n ∑ ∑
P Ei = P(Ei ) − P(Ei ∩ Ej ) + P(Ei ∩ Ej ∩ Ek )
i=1 i=1 i<j i<j<k

∑ ( ∩
ℓ )
· · · + (−1) ℓ+1
P Eim + · · ·
i1 <i2 <···<iℓ m=1

• Very difficult to calculate, often rely on the union bound:

(∪
n ) ∑
n
P Ei ≤ P(Ei ) .
i=1 i=1
Probability Review: Events

Independence:
• Two events E1 and E2 are independent if

P(E1 ∩ E2 ) = P(E1 )P(E2 ) .

Probability Review: Events

Independence:
• Two events E1 and E2 are independent if

P(E1 ∩ E2 ) = P(E1 )P(E2 ) .

• The events E1 , . . . , En are mutually independent (or just

independent) if, for all subsets I ⊆ {1, . . . , n},
(∩ ) ∏
P Ei = P(Ei ) .
i∈I i∈I
Probability Review: Events

Conditional Probability:
• The conditional probability that event E1 occurs given that event E2
occurs is
P(E1 ∩ E2 )
P(E1 |E2 ) = .
P(E2 )
Note that this is well-defined only if P(E2 ) > 0.
Probability Review: Events

Conditional Probability:
• The conditional probability that event E1 occurs given that event E2
occurs is
P(E1 ∩ E2 )
P(E1 |E2 ) = .
P(E2 )
Note that this is well-defined only if P(E2 ) > 0.

• Notice that if E1 and E2 are independent and P(E2 ) > 0,

P(E1 ∩ E2 ) P(E1 )P(E2 )

P(E1 |E2 ) = = = P(E1 ) .
P(E2 ) P(E2 )
Probability Review: Events

Law of Total Probability:

∞
∪
• If E1 , E2 , . . . are disjoint events such that Ω = Ei , then for any
i=1
event A
∞
∑ ∞
∑
P(A) = P(A ∩ Ei ) = P(A|Ei )P(Ei ) .
i=1 i=1

Bayes’ Law:
∞
∪
• If E1 , E2 , . . . are disjoint events such that Ω = Ei , then for any
i=1
event A
P(A|Ej )P(Ej )
P(Ej |A) = ∞ .
∑
P(A|Ei )P(Ei )
i=1
Probability Review: Random Variables

• A random variable X on a sample space Ω is a real-valued function,

X : Ω → R.
• Cumulative Distribution Function (cdf): FX (x) = P(X ≤ x).
Discrete Random Variables:
Probability Review: Random Variables

• A random variable X on a sample space Ω is a real-valued function,

X : Ω → R.
• Cumulative Distribution Function (cdf): FX (x) = P(X ≤ x).
Discrete Random Variables:
• X is discrete if it only takes values on a countable subset X of R.
• Probability Mass Function (pmf): For discrete random variables, we
can define the pmf pX (x) = P(X = x).
• Example 1: Bernoulli with parameter q.
{
1−q x=0
pX (x) =
q x=1.

• Example 2: Binomial with parameters n and q.

( )
n k
pX (k) = q (1 − q)n−k k = 0, 1, . . . , n
k
Probability Review: Random Variables

Continuous Random Variables:

• A random variable X is called continuous if there exists a
nonnegative function fX (x) such that
∫ b
P(a < X ≤ b) = fX (x)dx for all − ∞ < a < b < ∞ .
a

This function fX (x) is called the probability density function (pdf)

of X.
Probability Review: Random Variables

• Example 1: Uniform with parameters a and b.

{
1
b−a a<x≤b
fX (x) =
0 otherwise.

• Example 2: Gaussian with parameters µ and σ 2 .

1 (x−µ)2
fX (x) = √ e− 2σ 2
2πσ 2
• Example 3: Exponential with parameter λ.
{
λe−λx x≥0
fX (x) =
0 x<0.
Probability Review: Random Variables

Expectation:
∑
• Discrete rvs: E[g(X)] = g(x)pX (x)
x∈X
∫ ∞
• Continuous rvs: E[g(X)] = g(x)fX (x)dx
−∞

Special Cases

Mean:
∑
• Discrete rvs: E[X] = xpX (x)
x∈X
∫ ∞
• Continuous rvs: E[X] = xfX (x)dx
−∞
Probability Review: Random Variables

Expectation:
∑
• Discrete rvs: E[g(X)] = g(x)pX (x)
x∈X
∫ ∞
• Continuous rvs: E[g(X)] = g(x)fX (x)dx
−∞

Special Cases

Mean:
∑
• Discrete rvs: E[X] = xpX (x)
x∈X
∫ ∞
• Continuous rvs: E[X] = xfX (x)dx
−∞
Bernoulli Binomial Uniform Gaussian Exponential
• a+b 1
p np 2 µ λ
Probability Review: Random Variables

Special Cases Continued

mth Moment:
∑
• Discrete rvs: E[X m ] = xm pX (x)
x∈X
∫ ∞
• Continuous rvs: E[X m ] = xm fX (x)dx
−∞

Variance:
[ ] ( )2
• Var(X) = E (X − E[X])2 = E[X 2 ] − E[X]
Probability Review: Random Variables

Special Cases Continued

mth Moment:
∑
• Discrete rvs: E[X m ] = xm pX (x)
x∈X
∫ ∞
• Continuous rvs: E[X m ] = xm fX (x)dx
−∞

Variance:
[ ] ( )2
• Var(X) = E (X − E[X])2 = E[X 2 ] − E[X]

Bernoulli Binomial Uniform Gaussian Exponential

• (b−a)2
p(1 − p) np(1 − p) 12 σ2 1
λ2
Probability Review: Collections of Random Variables

Pairs of Random Variables (X, Y ):

• Joint cdf: FXY (x, y) = P(X ≤ x, Y ≤ y)
• Joint pmf: pXY (x, y) = P(X = x, Y = y) (for discrete rvs)
• Joint pdf: If fXY satisfies
∫ b∫ d
P(a < X ≤ b, c < Y ≤ d) = fXY (x, y)dydx
a c

for all −∞ < a < b < ∞ and −∞ < c < d < ∞ then fXY is called
the joint pdf of (X, Y ). (for continuous rvs)
• Marginalization: ∑
pY (y) = pXY (x, y)
x∈X
∫ ∞
fY (y) = fXY (x, y)dx
−∞
Probability Review: Collections of Random Variables

n-tuples of Random Variables (X1 , . . . , Xn ):

• Joint cdf: FX1 ···Xn (x1 , . . . , xn ) = P(X1 ≤ x1 , . . . , Xn ≤ xn )
• Joint pmf: pX1 ···Xn (x1 , . . . , xn ) = P(X1 = x1 , . . . , Xn = xn )
• Joint pdf: If fX1 ···Xn satisfies

P(a1 < X1 ≤ b1 , . . . , an < Xn ≤ bn )

∫ b1 ∫ bn
= ··· fX1 ···Xn (x1 , . . . , xn )dxn · · · dx1
a1 an

for all −∞ < ai < bi < ∞ then fX1 ···Xn is called the joint pdf of
(X1 , . . . , Xn ).
Probability Review: Collections of Random Variables

Independence of Random Variables:

• X1 , . . . , Xn are independent if
FX1 ···Xn (x1 , . . . , xn ) = FX1 (x1 ) · · · FXn (xn )
∀x1 , x2 , . . . , xn
• Equivalently, we can just check if

pX1 ···Xn (x1 , . . . , xn ) = pX1 (x1 ) · · · pXn (xn ) (discrete rvs)

fX1 ···Xn (x1 , . . . , xn ) = fX1 (x1 ) · · · fXn (xn ) (continuous rvs)

Probability Review: Collections of Random Variables

Conditional Probability Densities:

• Given discrete rvs X and Y with joint pmf pXY (x, y), the
conditional pmf of X given Y = y is defined to be
{ p (x,y)
XY
pY (y) pY (y) > 0
pX|Y (x|y) =
0 otherwise.

• Given continous rvs X and Y with joint pdf fXY (x, y), the
conditional pdf of X given Y = y is defined to be
{ f (x,y)
XY
fY (y) fY (y) > 0
fX|Y (x|y) =
0 otherwise.

• Note that if X and Y are independent, then pX|Y (x|y) = pX (x) or

fX|Y (x|y) = fX (x).
Probability Review: Collections of Random Variables

Linearity of Expectation:
• E[a1 X1 + · · · + an Xn ] = a1 E[X1 ] + · · · + an E[Xn ] even if the Xi
are dependent.

Expectation of Products:
• If X1 , . . . , Xn are independent, then
E[g1 (X1 ) · · · gn (Xn )] = E[g1 (X1 )] · · · E[gn (Xn )] for any
deterministic functions gi .
Probability Review: Collections of Random Variables

Conditional Expectation:
∑
• Discrete rvs: E[g(X)|Y = y] = g(x)pX|Y (x|y)
x∈X
∫ ∞
• Continuous rvs: E[g(X)|Y = y] = g(x)fX|Y (x|y)dx
−∞
Probability Review: Collections of Random Variables

Conditional Expectation:
∑
• Discrete rvs: E[g(X)|Y = y] = g(x)pX|Y (x|y)
x∈X
∫ ∞
• Continuous rvs: E[g(X)|Y = y] = g(x)fX|Y (x|y)dx
−∞

• E[Y |X = x] is a number. This number can be interpreted as a

function of x.
• E[Y |X] is a random variable. It is in fact a function of the random
variable X. (Note: A function of a random variable is a random
variable.)
[ ]
• Lemma: EX E[Y |X] = E[Y ].
Probability Review: Collections of Random Variables

Conditional Independence:
• X and Y are conditionally independent given Z if

pXY |Z (x, y|z) = pX|Z (x|z)pY |Z (y|z) (discrete rvs)

fXY |Z (x, y|z) = fX|Z (x|z)fY |Z (y|z) (continuous rvs)
Probability Review: Collections of Random Variables

Markov Chains:
• Random variables X, Y , and Z are said to form a Markov chain
X → Y → Z if the conditional distribution of Z depends only on Y
and is conditionally independent of X.
Probability Review: Collections of Random Variables

Markov Chains:
• Random variables X, Y , and Z are said to form a Markov chain
X → Y → Z if the conditional distribution of Z depends only on Y
and is conditionally independent of X.
• Specifically, the joint pmf (or pdf) can be factored as

pXY Z (x, y, z) = pX (x)pY |X (y|x)pZ|Y (z|y) (discrete rvs)

fXY Z (x, y, z) = fX (x)fY |X (y|x)fZ|Y (z|y) (continuous rvs) .
Probability Review: Collections of Random Variables

Markov Chains:
• Random variables X, Y , and Z are said to form a Markov chain
X → Y → Z if the conditional distribution of Z depends only on Y
and is conditionally independent of X.
• Specifically, the joint pmf (or pdf) can be factored as

pXY Z (x, y, z) = pX (x)pY |X (y|x)pZ|Y (z|y) (discrete rvs)

fXY Z (x, y, z) = fX (x)fY |X (y|x)fZ|Y (z|y) (continuous rvs) .

• X → Y → Z if and only if X and Z are conditionally independent

given Y .
• X → Y → Z implies Z → Y → X (and vice versa).

• If Z is a deterministic function of Y , i.e. Z = g(Y ), then

X → Y → Z automatically.
Probability Review: Inequalities

Convexity:
• A set X ⊆ Rn is convex if, for every x1 , x2 ∈ X and λ ∈ [0, 1], we
have that λx1 + (1 − λ)x2 ∈ X .
Probability Review: Inequalities

Convexity:
• A set X ⊆ Rn is convex if, for every x1 , x2 ∈ X and λ ∈ [0, 1], we
have that λx1 + (1 − λ)x2 ∈ X .
• A function g on a convex set X is convex if, for every x1 , x2 ∈ X
and λ ∈ [0, 1], we have that

g(λx1 + (1 − λ)x2 ) ≤ λg(x1 ) + (1 − λ)g(x2 ) .

Probability Review: Inequalities

Convexity:
• A set X ⊆ Rn is convex if, for every x1 , x2 ∈ X and λ ∈ [0, 1], we
have that λx1 + (1 − λ)x2 ∈ X .
• A function g on a convex set X is convex if, for every x1 , x2 ∈ X
and λ ∈ [0, 1], we have that

g(λx1 + (1 − λ)x2 ) ≤ λg(x1 ) + (1 − λ)g(x2 ) .

• A function g is concave if −g is convex.

Probability Review: Inequalities

Convexity:
• A set X ⊆ Rn is convex if, for every x1 , x2 ∈ X and λ ∈ [0, 1], we
have that λx1 + (1 − λ)x2 ∈ X .
• A function g on a convex set X is convex if, for every x1 , x2 ∈ X
and λ ∈ [0, 1], we have that

g(λx1 + (1 − λ)x2 ) ≤ λg(x1 ) + (1 − λ)g(x2 ) .

• A function g is concave if −g is convex.

Jensen’s Inequality:
• If g is a convex function and X is a random variable, then
( [ ]
g E[X]) ≤ E g(X)
Probability Review: Inequalities

Markov’s Inequality:
• Let X be a non-negative random variable. For any t > 0,

E[X]
P(X ≥ t) ≤ .
t
Probability Review: Inequalities

Markov’s Inequality:
• Let X be a non-negative random variable. For any t > 0,

E[X]
P(X ≥ t) ≤ .
t

Chebyshev’s Inequality:
• Let X be a random variable. For any ϵ > 0,
(  ) Var(X)
P X − E[X] > ϵ ≤ .
ϵ2

(You will prove these in Homework 1.)

Probability Review: Inequalities

Weak Law of Large Numbers (WLLN):

• Let Xi be a sequence of independent and identically distributed
(i.i.d.) random variables with finite mean, µ = E[Xi ] < ∞.

1∑
n
• Define the sample mean X̄n = Xi .
n
i=1

• For any ϵ > 0, the WLLN implies that

(  )
lim P X̄n − µ > ϵ = 0 .
n→∞

• That is, the sample mean converges (in probability) to the true
mean.
Probability Review: Inequalities

Strong Law of Large Numbers (SLLN):

• Let Xi be a sequence of independent and identically distributed
(i.i.d.) random variables with finite mean, µ = E[Xi ] < ∞.

1∑
n
• Define the sample mean X̄n = Xi .
n
i=1

• The SLLN implies that

({ })
P lim X̄n = µ =1.
n→∞

• That is, the sample mean converges (almost surely) to the true
mean.