APPPHYS202 - Tuesday 10 January 2012

Quantum(noncommutative)
probability
Quantumphysics
&informationtheory
Classical(commutative)
probability
Classicalphysics
&informationtheory

Given a (probabilistic) model and some data, what can we infer?

In the introductory part of the course we will restrict our attenion to finite discrete probability models, in both the
classical and quantum contexts. This will greatly simplify the mathematical and notational overhead. For those who
are interested, some additional background/reading on classical probability may be found the following
freely-available documents (URLs valid as of January 2012):
Introduction to Probability - Charles M. Grinstead and J. Laurie Snell
http://www.dartmouth.edu/~chance/teaching_aids/books_articles/probability_book/book.
ACM217 notes: Stochastic Calculus, Filtering and Stochastic Control - Ramon van Handel
http://www.princeton.edu/~rvan/acm217/acm217.html
Some additional material on noncommutative probability and the structure of quantum mechanics can be found in
some published books:
Lectures on Quantum Theory: Mathematical and Structural Foundations - C. J. Isham (World Sci)
Statistical Dynamics: A Stochastic Approach to Nonequilibrium Thermodynamics - R. F. Streater (Imperial
College Press)
see http://www.mth.kcl.ac.uk/~streater/bookcont.html for some known typos in the first
edition.
Review of classical probability/inference, with a view towards quantum
1. Discrete random variables as functions on a sample space
2. Probability distribution functions
1
3. Events
4. Algebras of random variables
5. Expectation, variance, and the notion of state
6. Matrix notation
7. Joint systems
8. Conditioning, Bayes Rule
9. Conditional expectation
Textbook discussions of basic probability often start with the example of rolling a six-sided die. Assuming the die and
its roller are deemed fair, we may assign equal probabilities to each of the six possible outcomes (number of spots
on the side that faces up). Before the die is cast, we can ask simple questions such as:
0 What is the probability that the result will be an even number?
0 What is the probability that the result will be either 1 or 2?
0 What is the probability that the result will not be 6?
All of you already know how to perform the simple calculations required to answer these correctly, so we will here
focus instead on using the example of a six-sided die to establish some formal terminology and concepts that will
help us eventually to understand the nature of the generalization from classical to quantum probability models. In
particular our goal for today will be to understand that a classical probability model comprises a sample space, an
algebra of random variables, and a probability state. We will also introduce a matrix notation for classical
observables that can naturally be generalized to accommodate quantum probablity models, and introduce
conditional expectation and conditional probability in this framework.
Random variables as functions on a sample space
First we note that a single six-sided die, once it has been rolled and come to rest on a level surface, has exactly six
possible configurations (also called outcomes) that are distinct and meaningful for our purposes. We will ignore the
exact spatial position of the die and its precise orientation, paying attention only to which face is up. Let us abstractly
identify the six possible outcomes resulting from a single die-roll with elements of the set

1
,
2
,
3
,
4
,
5
,
6
),
where the subscript index corresponds to the number of spots on the side that finally faces up. The set of all possible
outcomes is known as the sample space, and is usually denoted . For reasons that we will discuss below, subsets
of are called events.
Having established the set of possible outcomes, we can now define random variables (also called observables)
to be functions on . If a given random variable takes values in a set S we call it an S-valued random variable; real-
or integer-valued random variables are simply called random variables. For example consider the following random
variables, specified first in terms of intuitive definitions and second as explicit functions on :
0 X(-) : number of spots on the side facing up,
X(
1
) 1, X(
2
) 2, X(
3
) 3, X(
4
) 4, X(
5
) 5, X(
6
) 6.
0 Y(-) : sum of the numbers of spots on the five sides not facing up,
Y(
1
) 20, Y(
2
) 19, Y(
3
) 18, Y(
4
) 17, Y(
5
) 16, Y(
6
) 15.
0 Z(-) : value of the smallest prime number larger than the number of spots on the side facing up,
Z(
1
) 2, Z(
2
) 3, Z(
3
) 5, Z(
4
) 5, Z(
5
) 7, Z(
6
) 7.
The notation here is meant to emphasize the view of random variables as functions; note that these functions are not
necessarily one-to-one. We will implicitly treat the one-to-one random variable X(-) as a special variable that
indicates the numerical value of the die-roll, as this conforms to gambling convention, although in principle Y(-) or
any other one-to-one random variable could play the same role.
Note that if we know the exact configuration of our system, we implicitly know the exact value of all observables.
Similarly, if we know the exact value of any one-to-one random variable we can infer the configuration and thus the
2
exact value of all other observables.
Probability distribution functions
Another special function on the set of outcomes is the probability distribution function, which we will denote m(-).
This special function is defined by
m(
i
) Pr(
i
), i.
Since we are assuming that this is a fair die-roll, we have m(
i
) 1/6 for all i. Generally speaking, for a probability
distribution function in any scenario we require
0 m(
i
) 1,

i
m(
i
) 1.
Occasionally we may have cause to consider unnormalized probability distributions such that
i
m(
i
) 1, and
in such cases it is understood that Pr(
i
) m(
i
)/
i
m(
i
).
Events
Our next step is to discuss the association of events (subsets of ) with yes-or-no questions about the outcome.
Suppose I roll the six-sided die but do not show you the result. Any relevant yes-or-no question you could ask me
regarding the outcome of the die-roll can be associated with a subset E , such that I will say yes if and only if the
actual outcome is in E. For example:
0 Was the result 1? E
1
)
0 Was the result an even number? E
2
,
4
,
6
)
0 Was the result 1 or 2? E
1
,
2
)
0 Was the result not 6? E
1
,
2
,
3
,
4
,
5
)
Of course there is more than one way to formulate a yes-or-no question corresponding to a given set of elements:
0 E
1
,
2
) : Was the result less than 3?
0 E
1
) : Was the result not an even number, and less than 3?
Clearly the information content of the answer to a yes-or-no question regarding membership in a given subset
depends only on the subset, and is independent of the precise way that the question is worded.
Note that knowledge of membership in a subset E implies knowledge of membership in the complementary
subset E
C
. In words, this corresponds to the fact that the answer to a yes-or-no question implies the answer to the
negation of this question. Likewise, if we have knowledge of membership in two different subsets E
1
and E
2
, then we
can infer membership in the combined subsets E
1
E
2
and E
1
E
2
(exercise: write out the corresponding truth
tables). It is useful to note in this context that
E
1
E
2
(E
1
C
E
2
C
)
C
,
which means that complementation and union are really the essential operations in this type of inference game. In
any case we note that, if I allow you to ask about membership in a starter collection of subsets E
i
), you can
actually infer membership in a larger collection of subsets generated by complementation and union - the -algebra
generated by E
i
) (for explicit definitions of set complement, union, intersection and difference see Grinstead and
Snell, p. 21-22). For example with our six-sided die, if our starter set is E
3
, E
even
) with
E
3

1
,
2
,
3
), E
even

2
,
4
,
6
),
if I tell you whether or not a given configuration

is in E
3
and whether or not it is in E
even
you can infer whether or
not it is in any of the following subsets:
3
E
odd
E
even
C

1
,
3
,
5
),
E
4
E
3
C

4
,
5
,
6
),

2
) E
even
E
3
,

5
) E
odd
E
4
,

1
,
3
) E
odd
E
3
,

4
,
6
) E
even
E
4
,
as well as any possible union of these subsets.
Note that random variables can also serve to define events:
0 Was the result of the die-roll such that X 5? E
5
)
0 Was the result of the die-roll such that Y 15 or Y 16? E
5
,
6
)
0 Was the result of the die-roll such that Z 5? E
3
,
4
)
Knowing the value of a random variable does not necessarily allow you to determine the exact configuration, but you
can narrow it down to a subset of . The term level set is commonly used to refer to the event that contains all
configurations for which a random variable assumes a given value. Note that the level sets of a random variable are
non-overlapping, and that the union of all level sets of a random variable is . There can of course be cases where
joint knowledge of the values of two random variables allows you to infer an exact configuration, even if knowing only
one of them would not be sufficient (this may remind you of the concept of a complete set of commuting observables
in quantum mechanics, but the analogy is rather subtle as we shall see).
It is natural to extend the probability distribution function m(-) so that it is defined not only on elementary
outcomes but also on events. Explicitly,
m(E)

iE
m(
i
).
When viewed as a function from subsets to the reals, m(-) is often referred to as a probability measure (especially in
scenarios with continuous random variables). It is easy to show that the following properties hold [Grinstead and
Snell, Theorem 1.1]:
1. m(E) 0 for every E .
2. m() 1.
3. If E F then m(E) m(F).
4. If A and B are disjoint subsets of , then m(A B) m(A) m(B).
5. m(A
C
) 1 m(A) for every A .
Here A
C
indicates the complement of A in , as in our above discussion of events.
Note that the probability distribution function thus induces probabilities for the values of random variables. For a
random variable A, if we define
E
A,a

i
: A(
i
) a)
as the event A a, then
Pr(A a) m(E
A,a
)

iE
A,a
m(
i
).
For example X 5 occurs only for
5
), so Pr(X 5) m(
5
) 1/6. On the other hand,
Pr(Z 5) m(
3
) m(
4
) 1/3.
Algebras of random variables
Once we have defined some random variables, such as X, Y, Z, it is very easy to generate more (here we will assume
that all random variables can be viewed as taking real values). Note that sums and products of random variables are
themselves random variables, as are the products of random variables with real numbers. Hence, random variables
have a natural algebraic structure. For example, if we define
R(-) X(-) Z(-),
4
with , real numbers, then
R(
1
) 2, R(
2
) 2 3, R(
3
) 3 5,
R(
4
) 4 5, R(
5
) 5 7, R(
6
) 6 7.
Similarly,
Z
2
(-) |Z(-)]
2
has values
Z
2
(
1
) 4, Z
2
(
2
) 9, Z
2
(
3
) 25, Z
2
(
4
) 25, Z
2
(
5
) 49, Z
2
(
6
) 49,
and
XZ(-) X(-)Z(-)
has values
XZ(
1
) 2, XZ(
2
) 6, XZ(
3
) 15, XZ(
4
) 20, XZ(
5
) 35, XZ(
6
) 42.
The probability distribution function on clearly provides probability distribution functions for such random variables
as well.
An indicator function
E
(-) of an event (subset) E is a random variable such that

E
(
i
) 1,
i
E,
0,
i
E.
Technically speaking, any random variable can be expressed in terms of indicator functions on its level sets:
R(-)

i
r
i

r
i
(-),
where R(-) takes values in the set r
i
) and
ri
is the level set corresponding to the value r
i
. For example,
Z(-) 2

1
)
(-) 3

2
)
(-) 5

3
,
4
)
(-) 7
5,
6
)
(-).
It thus appears that indicator functions are like basis functions for random variables. Note that for two events A and
B,

AB
(
i
)
A
(
i
)
B
(
i
).
Hence for a pair of random variables R(-) and T(-),
R(-)T(-)

i
r
i

r
i
(-)

j
t
j

t
j
(-)

i, j
r
i
t
j

r
i
t
j
(-) T(-)R(-).
Expectation, variance, and the notion of state
The expectation of a random variable R(-), which we will write R), is defined as
R)

R(
i
)m(
i
).
This is the average, or mean value of R with respect to the probability distribution function m(-). Note that for
indicator functions,

E) m(E).
Similarly, the variance of R(-) is defined as
var|R] R
2
)

i
R
2
(
i
)m(
i
)

i
|R(
i
)]
2
m(
i
).
It is common also to define the standard deviation of R(-), also called the uncertainty of R(-), as
std|R] R
2
) R)
2

|R(
i
)]
2
m(
i
)

R(
i
)m(
i
)
2
.
It is common also to define the covariance of two random variables A(-) and B(-) as
5
cov|A, B] (A A))(B B))) AB) A)B)

A(
i
)B(
i
)m(
i
)

A(
i
)m(
i
)

B(
i
)m(
i
) .
It should be clear from these definitions that, in general, R
2
) R)
2
and AB) A)B). If cov|A, B] 0 we say that
A(-) and B(-) are linearly independent random variables.
Formally, a state is a consistent assignment of an expectation value to every random variable in an algebra. It
should be clear from the above that a state specifies variances and covariances by virtue of the fact that if A(-) and
B(-) are random variables in our algebra, then so are A
2
(-), B
2
(-) and AB(-). The probability measure m(-) is a
compact way of summarizing the state on an algebra of random variables. Note that state and configuration are quite
different in our useage of the terms - classically we assume that there exists an actual configuration of the system in
question (the actual disposition of the die after it has been rolled), which may or may not be known to anyone, but we
also have a state of knowledge/belief that summarizes the information we use to make predictions within a
probabilistic framework.
Matrix notation
In a finite discrete setting, for which the sample space contains N elements, it is natural to associate random
variables with N N real matrices. For an arbitrary random variable R(-), we simply place the values R(
i
) along the
diagonal and put zeros everywhere else. Hence, continuing with our example of the six-sided die:
X(-)
1 0 0 0 0 0
0 2 0 0 0 0
0 0 3 0 0 0
0 0 0 4 0 0
0 0 0 0 5 0
0 0 0 0 0 6
, Z(-)
2 0 0 0 0 0
0 3 0 0 0 0
0 0 5 0 0 0
0 0 0 5 0 0
0 0 0 0 7 0
0 0 0 0 0 7
.
We use (X) to denote the matrix representation of a random variable X(-). With a bit of thought you can convince
yourself that with this matrix representation, we can use the usual rules of matrix arithmetic and multiplication to
carry out algebraic manipulations among random variables. For example,
R(-) X(-) Z(-)
1 0 0 0 0 0
0 2 0 0 0 0
0 0 3 0 0 0
0 0 0 4 0 0
0 0 0 0 5 0
0 0 0 0 0 6

2 0 0 0 0 0
0 3 0 0 0 0
0 0 5 0 0 0
0 0 0 5 0 0
0 0 0 0 7 0
0 0 0 0 0 7
diag( 2, 2 3, 3 5, 4 5, 5 7, 6 7),
where the diag() notation hopefully is obvious. Note that because of the fact that all matrices we use in this
classical probability setting are diagonal, the matrix representations of an algebra of random variables form a
commutative matrix algebra.
We note that the probability distribution can be written in exactly the same matrix notation, and that we thus
arrive with the convenient expressions such as
6
X)

X(
i
)m(
i
)
Tr
1 0 0 0 0 0
0 2 0 0 0 0
0 0 3 0 0 0
0 0 0 4 0 0
0 0 0 0 5 0
0 0 0 0 0 6
1/6 0 0 0 0 0
0 1/6 0 0 0 0
0 0 1/6 0 0 0
0 0 0 1/6 0 0
0 0 0 0 1/6 0
0 0 0 0 0 1/6
.
We will use the suggestive notation diag(m(
1
), , m(
N
)) for the matrix representing the probability distribution
function. Hence, in general, the expectation R) of an arbitrary random variable R can be computed by taking the
trace of the product of with (R). The matrix provides a convenient representation of a state for our algebra of
random variables.
Indicator functions have a somewhat special appearance in this matrix notation, as they correspond to matrices
with zeros and ones on the diagonal. Viewed as linear operators, they are therefore projection (idempotent)
operators. For example, the indicator function
E
(-) for the event E
1
,
2
) has matrix representation

E
(-)
1 0 0 0 0 0
0 1 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
,
where clearly (
E
)
2
(
E
). It should be evident that the matrix representations of the indicator functions on all of the
individual outcomes
1
),
2
), ,
N
) provide a linear basis for the commutative matrix algebra representing all
possible random variables on . In particular,
(R)

i1
N
R(
i
)(
i )
).
It may occur to you that this is actually a sort of spectral decomposition of R viewed as a linear operator. Hopefully,
this perspective also highlights the fact that we can easily identify sub-algebras. For example if we think about the
linear span of the matrix representations of indicator functions on
1
,
3
,
5
) and
2
,
4
,
6
), we obtain a closed
matrix algebra for which the first, third and fifth diagonal elements are always the same, as are the second, fourth
and sixth. It is only really two-dimensional.
Note that once we have obtained the matrix representations for the observables that we care about, and for the
state, we can actually forget about and the underlying configurations! Our original notion of random variables as
functions on a sample space dictated the dimension of the matrix representations and their diagonality (required for
multiplication to be commutative).
Joint systems
Suppose now we have two six-sided dice. We can at first consider them to be independent systems living in
separate sample spaces:

A

1
A
,
2
A
,
3
A
,
4
A
,
5
A
,
6
A
),
B

1
B
,
2
B
,
3
B
,
4
B
,
5
B
,
6
B
).
Clearly we can define random variables on each space, such as
X
A
(
i
A
) i, X
B
(
j
B
) j.
Note that at this level of description,
1
B
is not in the domain of X
A
(-) and therefore X
A
(
i
B
) is undefined. Likewise, we
7
have two probability distribution functions m
A
(-) and m
B
(-), which we might as well take to be uniform.
We can clearly construct a joint sample space by taking Cartesian products:

AB

1
A

1
B
,
1
A

2
B
,
1
A

3
B
,
1
A

4
B
,
1
A

5
B
,
1
A

6
B
,
2
A

1
B
, ,
6
A

6
B
).
Now
AB
has 36 elements, corresponding to all possible outcomes of the rolling of a pair of six-sided dice. What
about the random variables and probability distribution functions? Consider the following definition:
R
AB
(-) R
A
(-) R
B
(-),
R
AB
(
i
A

j
B
) R
A
(
i
A
)R
B
(
j
B
),
where the final expression indicates simple scalar multiplication of the numerical values of R
A
(
i
A
) and R
B
(
j
B
).
Making use of the identity functions
1
A
(-)

A (-), 1
B
(-)

B(-),
we can thus define ampliations of the random variables we initially define on the factor spaces
A
and
B
to the joint
space
AB
. For example,
X
A
(-)
B
|X
A
](-) X
A
(-) 1
B
(-),
B
|X
A
](
i
A

j
B
) X
A
(
i
A
) i,
X
B
(-)
A
|X
B
](-) 1
A
(-) X
B
(-),
A
|X
B
](
i
A

j
B
) X
B
(
j
B
) j.
Often we will simply write
X
A
(
i
A

j
B
) i, X
B
(
i
A

j
B
) j,
with all the ampliation stuff implied. Note that we can now also consider things like
X
AB
(-) X
A
(-) X
B
(-), X
AB
(
i
A

j
B
) ij,
X
AB
(-) X
A
(-) 1
B
(-) 1
A
(-) X
B
(-), X
AB
(
i
A

j
B
) i j.
Normally in games we consider X
AB
(-) to correspond to the numerical value of the roll. Incidentally, note that X
AB
(-)
is a random variable on
AB
that does not simply factor into the product of a random variable on
A
with another
random variable on
B
.
Turning now to the probability distribution functions, we note that
m
AB
(-) m
A
(-) m
B
(-), m
AB
(
i
A

j
B
) m
A
(
i
A
)m
B
(
j
B
) 1/36
provides the proper joint probability distribution function on
AB
(it is clearly normalized). While we are free to define
m
AB
(-) m
A
(-) 1
B
(-) 1
A
(-) m
B
(-) 1/3,
this function on
AB
is not a valid probability measure. Note that the action of m
AB
(-) on subsets of
AB
follows in an
obvious way.
While the matrix representations of our new joint random variables are rather cumbersome to write down, we
note that they have dimension 36 which is the product of the matrix representation dimensions on the factor spaces.
In fact, using notation familiar from quantum mechanics we can write
(R
AB
) (R
A
) (R
B
),
where denotes tensor (Kronecker) product. If you didnt see this in your prevous quantum class dont worry - well
review this later. If you do know how to take tensor products of matrices, perhaps you could verify the above relation
for (X
AB
) and (X
AB
).
Finally we mention the issue of marginalization. Suppose we retain the joint sample space
AB
but I now tell you
that the dice are weighted and that I am going to roll them in some sneaky way that could correlate their outcomes. I
summarize the information numerically by giving you a new joint probability distribution function n
AB
(-). If we forget
about the B die, what is the marginal probability distribution function n
A
(-) for the A die only? In the functional
notation we can write
n
A
(
i
A
)

j1
6
n
AB
(
i
A

j
B
).
In matrix notation we would like to have a procedure for going from diag(n
AB
(
1
A

1
B
), , n
AB
(
6
A

6
B
)) to
diag(n
A
(
1
A
), , n
A
(
6
A
)) via linear algebra-type operations. Again, from previous quantum classes it may not surprise
you to hear that this is a partial trace operation; this also will be reviewed a bit later on in the course.
8
Conditioning
Suppose I roll the dice without showing you the exact outcome, but I tell you that X
AB
7. There are obviously
several joint configurations consistent with this but we can rule out others, such as
1
A

1
B
; how should you update
your original m
AB
(-) to obtain a conditional probability distribution m
AB
(- | X
AB
7)?
Most of you will have seen Bayes Rule on some previous occasion:
Pr(E| F)
Pr(F| E) Pr(E)
Pr(F)
,
which can be thought of as a summary of the equations
Pr(E, F) Pr(E| F) Pr(F),
Pr(F, E) Pr(F| E) Pr(E),
Pr(E, F) Pr(F, E).
For the present discussion it will be most useful to use the slightly modified form
Pr(E| F)
Pr(E, F)
Pr(F)
.
Here Pr(E, F) is the joint probability of E and F, while Pr(E| F) is the conditional probability of E given F. The
probabilitiles Pr(E) and Pr(F) are understood to be prior probabilities, that is, the probabilites we would have
assigned to the events E and F before gaining any updated information. Note here the use of the term events, which
should immediately alert you to how we are going to proceed. Inovking Bayes Rule for our dice scenario, we define
E
i
A

j
B
),
F
AB
: X
AB
() 7),
(F is a level set of X
AB
(-)) and find
m
AB
(
i
A

j
B
| X
AB
7) Pr(E| F)
Pr(E, F)
Pr(F)

m
AB
(E F)
m
AB
(F)
.
Clearly the numerator vanishes for any configuration not in the level set, and is equal to 1/36 for any configuration
that is in the level set. The denominator is actually independent of
i
A

j
B
, and in fact can be seen to be equal to the
sum over all
AB
of m
AB
( F). Hence it is simple a normalization factor for the conditional probability
distribution. So in the end,
m
AB
(
i
A

j
B
| X
AB
7)
1
6
,
i
A

j
B
F,
0,
i
A

j
B
F,
F
1
A

6
B
,
2
A

5
B
,
3
A

4
B
,
4
A

3
B
,
5
A

2
B
,
6
A

1
B
).
The basic structure of Bayes Rule, which we have just highlighted, is that you condition your probability distribution
function by eliminating all configurations that are inconsistent with the information gained and then renormalizing
whatever is left over.
9