Conditional expectation

In probability theory, the conditional expectation, conditional expected value, or conditional mean of a random variable is its expected value evaluated with respect to the conditional probability distribution. If the random variable can take on only a finite number of values, the "conditions" are that the variable can only take on a subset of those values. More formally, in the case when the random variable is defined over a discrete probability space, the "conditions" are a partition of this probability space.

Depending on the context, the conditional expectation can be either a random variable or a function. The random variable is denoted analogously to conditional probability. The function form is either denoted or a separate function symbol such as is introduced with the meaning .

Examples

edit

Example 1: Dice rolling

edit

Consider the roll of a fair ‹See Tfd›die and let A = 1 if the number is even (i.e., 2, 4, or 6) and A = 0 otherwise. Furthermore, let B = 1 if the number is prime (i.e., 2, 3, or 5) and B = 0 otherwise.

1 2 3 4 5 6
A 0 1 0 1 0 1
B 0 1 1 0 1 0

The unconditional expectation of A is  , but the expectation of A conditional on B = 1 (i.e., conditional on the die roll being 2, 3, or 5) is  , and the expectation of A conditional on B = 0 (i.e., conditional on the die roll being 1, 4, or 6) is  . Likewise, the expectation of B conditional on A = 1 is  , and the expectation of B conditional on A = 0 is  .

Example 2: Rainfall data

edit

Suppose we have daily rainfall data (mm of rain each day) collected by a weather station on every day of the ten-year (3652-day) period from January 1, 1990, to December 31, 1999. The unconditional expectation of rainfall for an unspecified day is the average of the rainfall amounts for those 3652 days. The conditional expectation of rainfall for an otherwise unspecified day known to be (conditional on being) in the month of March, is the average of daily rainfall over all 310 days of the ten–year period that fall in March. Similarly, the conditional expectation of rainfall conditional on days dated March 2 is the average of the rainfall amounts that occurred on the ten days with that specific date.

History

edit

The related concept of conditional probability dates back at least to Laplace, who calculated conditional distributions. It was Andrey Kolmogorov who, in 1933, formalized it using the Radon–Nikodym theorem.[1] In works of Paul Halmos[2] and Joseph L. Doob[3] from 1953, conditional expectation was generalized to its modern definition using sub-σ-algebras.[4]

Definitions

edit

Conditioning on an event

edit

If A is an event in   with nonzero probability, and X is a discrete random variable, the conditional expectation of X given A is

 

where the sum is taken over all possible outcomes of X.

If  , the conditional expectation is undefined due to the division by zero.

Discrete random variables

edit

If X and Y are discrete random variables, the conditional expectation of X given Y is

 

where   is the joint probability mass function of X and Y. The sum is taken over all possible outcomes of X.

Remark that as above the expression is undefined if  .

Conditioning on a discrete random variable is the same as conditioning on the corresponding event:

 

where A is the set  .

Continuous random variables

edit

Let   and   be continuous random variables with joint density    's density   and conditional density   of   given the event   The conditional expectation of   given   is

 

When the denominator is zero, the expression is undefined.

Conditioning on a continuous random variable is not the same as conditioning on the event   as it was in the discrete case. For a discussion, see Conditioning on an event of probability zero. Not respecting this distinction can lead to contradictory conclusions as illustrated by the Borel-Kolmogorov paradox.

L2 random variables

edit

All random variables in this section are assumed to be in  , that is square integrable. In its full generality, conditional expectation is developed without this assumption, see below under Conditional expectation with respect to a sub-σ-algebra. The   theory is, however, considered more intuitive[5] and admits important generalizations. In the context of   random variables, conditional expectation is also called regression.

In what follows let   be a probability space, and   in   with mean   and variance  . The expectation   minimizes the mean squared error:

 

The conditional expectation of X is defined analogously, except instead of a single number  , the result will be a function  . Let   be a random vector. The conditional expectation   is a measurable function such that

 

Note that unlike  , the conditional expectation   is not generally unique: there may be multiple minimizers of the mean squared error.

Uniqueness

edit

Example 1: Consider the case where Y is the constant random variable that is always 1. Then the mean squared error is minimized by any function of the form

 

Example 2: Consider the case where Y is the 2-dimensional random vector  . Then clearly

 

but in terms of functions it can be expressed as   or   or infinitely many other ways. In the context of linear regression, this lack of uniqueness is called multicollinearity.

Conditional expectation is unique up to a set of measure zero in  . The measure used is the pushforward measure induced by Y.

In the first example, the pushforward measure is a Dirac distribution at 1. In the second it is concentrated on the "diagonal"  , so that any set not intersecting it has measure 0.

Existence

edit

The existence of a minimizer for   is non-trivial. It can be shown that

 

is a closed subspace of the Hilbert space  .[6] By the Hilbert projection theorem, the necessary and sufficient condition for   to be a minimizer is that for all   in M we have

 

In words, this equation says that the residual   is orthogonal to the space M of all functions of Y. This orthogonality condition, applied to the indicator functions  , is used below to extend conditional expectation to the case that X and Y are not necessarily in  .

Connections to regression

edit

The conditional expectation is often approximated in applied mathematics and statistics due to the difficulties in analytically calculating it, and for interpolation.[7]

The Hilbert subspace

 

defined above is replaced with subsets thereof by restricting the functional form of g, rather than allowing any measurable function. Examples of this are decision tree regression when g is required to be a simple function, linear regression when g is required to be affine, etc.

These generalizations of conditional expectation come at the cost of many of its properties no longer holding. For example, let M be the space of all linear functions of Y and let   denote this generalized conditional expectation/  projection. If   does not contain the constant functions, the tower property   will not hold.

An important special case is when X and Y are jointly normally distributed. In this case it can be shown that the conditional expectation is equivalent to linear regression:

 

for coefficients   described in Multivariate normal distribution#Conditional distributions.

Conditional expectation with respect to a sub-σ-algebra

edit
 
Conditional expectation with respect to a σ-algebra: in this example the probability space   is the [0,1] interval with the Lebesgue measure. We define the following σ-algebras:  ;   is the σ-algebra generated by the intervals with end-points 0, 14, 12, 34, 1; and   is the σ-algebra generated by the intervals with end-points 0, 12, 1. Here the conditional expectation is effectively the average over the minimal sets of the σ-algebra.

Consider the following:

  •   is a probability space.
  •   is a random variable on that probability space with finite expectation.
  •   is a sub-σ-algebra of  .

Since   is a sub  -algebra of  , the function   is usually not  -measurable, thus the existence of the integrals of the form  , where   and   is the restriction of   to  , cannot be stated in general. However, the local averages   can be recovered in   with the help of the conditional expectation.

A conditional expectation of X given  , denoted as  , is any  -measurable function   which satisfies:

 

for each  .[8]

As noted in the   discussion, this condition is equivalent to saying that the residual   is orthogonal to the indicator functions  :

 

Existence

edit

The existence of   can be established by noting that   for   is a finite measure on   that is absolutely continuous with respect to  . If   is the natural injection from   to  , then   is the restriction of   to   and   is the restriction of   to  . Furthermore,   is absolutely continuous with respect to  , because the condition

 

implies

 

Thus, we have

 

where the derivatives are Radon–Nikodym derivatives of measures.

Conditional expectation with respect to a random variable

edit

Consider, in addition to the above,

  • A measurable space  , and
  • A random variable  .

The conditional expectation of X given Y is defined by applying the above construction on the σ-algebra generated by Y:

 

By the Doob–Dynkin lemma, there exists a function   such that

 

Discussion

edit
  • This is not a constructive definition; we are merely given the required property that a conditional expectation must satisfy.
    • The definition of   may resemble that of   for an event   but these are very different objects. The former is a  -measurable function  , while the latter is an element of   and   for  .
    • Uniqueness can be shown to be almost sure: that is, versions of the same conditional expectation will only differ on a set of probability zero.
  • The σ-algebra   controls the "granularity" of the conditioning. A conditional expectation   over a finer (larger) σ-algebra   retains information about the probabilities of a larger class of events. A conditional expectation over a coarser (smaller) σ-algebra averages over more events.

Conditional probability

edit

For a Borel subset B in  , one can consider the collection of random variables

 

It can be shown that they form a Markov kernel, that is, for almost all  ,   is a probability measure.[9]

The Law of the unconscious statistician is then

 

This shows that conditional expectations are, like their unconditional counterparts, integrations, against a conditional measure.

General Definition

edit

In full generality, consider:

  • A probability space  .
  • A Banach space  .
  • A Bochner integrable random variable  .
  • A sub-σ-algebra  .

The conditional expectation of   given   is the up to a  -nullset unique and integrable  -valued  -measurable random variable   satisfying

 

for all  .[10][11]

In this setting the conditional expectation is sometimes also denoted in operator notation as  .

Basic properties

edit

All the following formulas are to be understood in an almost sure sense. The σ-algebra   could be replaced by a random variable  , i.e.  .

  • Pulling out independent factors:
    • If   is independent of  , then  .
Proof

Let  . Then   is independent of  , so we get that

 

Thus the definition of conditional expectation is satisfied by the constant random variable  , as desired.  

    • If   is independent of  , then  . Note that this is not necessarily the case if   is only independent of   and of  .
    • If   are independent,   are independent,   is independent of   and   is independent of  , then  .
  • Stability:
    • If   is  -measurable, then  .
Proof

For each   we have  , or equivalently

 

Since this is true for each  , and both   and   are  -measurable (the former property holds by definition; the latter property is key here), from this one can show

 

And this implies   almost everywhere.  

    • In particular, for sub-σ-algebras   we have  . (Note this is different from the tower property below.)
    • If Z is a random variable, then  . In its simplest form, this says  .
  • Pulling out known factors:
    • If   is  -measurable, then  .
Proof

All random variables here are assumed without loss of generality to be non-negative. The general case can be treated with  .

Fix   and let  . Then for any  

 

Hence   almost everywhere.

Any simple function is a finite linear combination of indicator functions. By linearity the above property holds for simple functions: if   is a simple function then  .

Now let   be  -measurable. Then there exists a sequence of simple functions   converging monotonically (here meaning  ) and pointwise to  . Consequently, for  , the sequence   converges monotonically and pointwise to  .

Also, since  , the sequence   converges monotonically and pointwise to  

Combining the special case proved for simple functions, the definition of conditional expectation, and deploying the monotone convergence theorem:

 

This holds for all  , whence   almost everywhere.  

    • If Z is a random variable, then  .
  • Law of total expectation:  .[12]
  • Tower property:
    • For sub-σ-algebras   we have  .
      • A special case   recovers the Law of total expectation:  .
      • A special case is when Z is a  -measurable random variable. Then   and thus  .
      • Doob martingale property: the above with   (which is  -measurable), and using also  , gives  .
    • For random variables   we have  .
    • For random variables   we have  .
  • Linearity: we have   and   for  .
  • Positivity: If   then  .
  • Monotonicity: If   then  .
  • Monotone convergence: If   then  .
  • Dominated convergence: If   and   with  , then  .
  • Fatou's lemma: If   then  .
  • Jensen's inequality: If   is a convex function, then  .
  • Conditional variance: Using the conditional expectation we can define, by analogy with the definition of the variance as the mean square deviation from the average, the conditional variance
    • Definition:  
    • Algebraic formula for the variance:  
    • Law of total variance:  .
  • Martingale convergence: For a random variable  , that has finite expectation, we have  , if either   is an increasing series of sub-σ-algebras and   or if   is a decreasing series of sub-σ-algebras and  .
  • Conditional expectation as  -projection: If   are in the Hilbert space of square-integrable real random variables (real random variables with finite second moment) then
    • for  -measurable  , we have  , i.e. the conditional expectation   is in the sense of the L2(P) scalar product the orthogonal projection from   to the linear subspace of  -measurable functions. (This allows to define and prove the existence of the conditional expectation based on the Hilbert projection theorem.)
    • the mapping   is self-adjoint:  
  • Conditioning is a contractive projection of Lp spaces  . I.e.,   for any p ≥ 1.
  • Doob's conditional independence property:[13] If   are conditionally independent given  , then   (equivalently,  ).

See also

edit

Probability laws

edit

Notes

edit
  1. ^ Kolmogorov, Andrey (1933). Grundbegriffe der Wahrscheinlichkeitsrechnung (in German). Berlin: Julius Springer. p. 46.
  2. ^ Oxtoby, J. C. (1953). "Review: Measure theory, by P. R. Halmos" (PDF). Bull. Amer. Math. Soc. 59 (1): 89–91. doi:10.1090/s0002-9904-1953-09662-8.
  3. ^ J. L. Doob (1953). Stochastic Processes. John Wiley & Sons. ISBN 0-471-52369-0.
  4. ^ Olav Kallenberg: Foundations of Modern Probability. 2. edition. Springer, New York 2002, ISBN 0-387-95313-2, p. 573.
  5. ^ "probability - Intuition behind Conditional Expectation". Mathematics Stack Exchange.
  6. ^ Brockwell, Peter J. (1991). Time series : theory and methods (2nd ed.). New York: Springer-Verlag. ISBN 978-1-4419-0320-4.
  7. ^ Hastie, Trevor (26 August 2009). The elements of statistical learning : data mining, inference, and prediction (PDF) (Second, corrected 7th printing ed.). New York. ISBN 978-0-387-84858-7.{{cite book}}: CS1 maint: location missing publisher (link)
  8. ^ Billingsley, Patrick (1995). "Section 34. Conditional Expectation". Probability and Measure (3rd ed.). John Wiley & Sons. p. 445. ISBN 0-471-00710-2.
  9. ^ Klenke, Achim (30 August 2013). Probability theory : a comprehensive course (Second ed.). London. ISBN 978-1-4471-5361-0.{{cite book}}: CS1 maint: location missing publisher (link)
  10. ^ Da Prato, Giuseppe; Zabczyk, Jerzy (2014). Stochastic Equations in Infinite Dimensions. Cambridge University Press. p. 26. doi:10.1017/CBO9781107295513. ISBN 978-1-107-05584-1. (Definition in separable Banach spaces)
  11. ^ Hytönen, Tuomas; van Neerven, Jan; Veraar, Mark; Weis, Lutz (2016). Analysis in Banach Spaces, Volume I: Martingales and Littlewood-Paley Theory. Springer Cham. doi:10.1007/978-3-319-48520-1. ISBN 978-3-319-48519-5. (Definition in general Banach spaces)
  12. ^ "Conditional expectation". www.statlect.com. Retrieved 2020-09-11.
  13. ^ Kallenberg, Olav (2001). Foundations of Modern Probability (2nd ed.). York, PA, USA: Springer. p. 110. ISBN 0-387-95313-2.

References

edit
  • William Feller, An Introduction to Probability Theory and its Applications, vol 1, 1950, page 223
  • Paul A. Meyer, Probability and Potentials, Blaisdell Publishing Co., 1966, page 28
  • Grimmett, Geoffrey; Stirzaker, David (2001). Probability and Random Processes (3rd ed.). Oxford University Press. ISBN 0-19-857222-0., pages 67–69
edit