Author 
Message 
Randy Poe science forum Guru
Joined: 24 Mar 2005
Posts: 2485

Posted: Wed Jul 19, 2006 9:34 pm Post subject:
Expectation value in terms of cumulative distribution



I have to estimate some means and variances from empirical
data, and for various reasons what I have to work with are
estimates of cumulative distributions rather than densities
Actually, what I have are quantile levels, a set of values
q(p) = x such that P(X<=x) = F(x) = p at evenly spaced values of p
from 0 to 1, which gives me in effect estimates of F(x) at unevenly
spaced intervals of x.
So I started wondering if I could estimate E[x] and E[x^2] directly
from F(x) and came up with an elementary theorem in about
30 seconds which I don't remember seeing. The proof is only
a couple of lines. Even though the proof seems pretty straightforward,
I'd like to know if this is in fact a theorem anyone has seen before.
Theorem: Let X be a random variable with finite support [a,b],
density function f(x) and cumulative probability F(x).
Then E[x] = b  integral(a,b)F(x) dx and
E[x^2] = b^2  2*integral(a,b) x F(x) dx
Proof: Integrate by parts.
E[x] = integral(a,b) x f(x) dx
Choose u = x, dv = f(x) dx => du = dx, v = F(x)
E[x] = (b*F(b)  a*F(a))  integral(a,b) F(x) dx which gives the first
result since F(a) = 0, F(b) = 1.
E[x^2] = integral(a,b) x^2 f(x) dx
= (b^2*F(b)  a^2*F(a))  integral(a,b) 2x F(x) dx
which gives the second result.
 Randy 

Back to top 


Stephen J. Herschkorn science forum Guru
Joined: 24 Mar 2005
Posts: 641

Posted: Wed Jul 19, 2006 9:45 pm Post subject:
Re: Expectation value in terms of cumulative distribution



It is wellknown that if X is a nonnegative random variable, then EX
= integral(x=0..infty, P{X > x}). The proof is easy to follow:
Let I(x) be the indicator variable for {X > x}. Then X =
int(x=0..infty, I(x)). Hence,
EX = E int(x=0..infty, I(x)) = int(x=0..infty, EI(x)) by Tonelli's
theorem. The integrand EI(x) = P{X > x}.
Note that X need not be continuous.
As a corollary, for nonnonegative X and a > 0,
E[X^a] = int(u=0..infty, P{X^a > u}) = int(u=0..infty, P{X >
u^(1/a)). Thus,
E[X^a] = a int(x=0..infty, x^(a1) P{X > x}) by change of variable in
the integral.

Stephen J. Herschkorn sjherschko@netscape.net
Math Tutor on the Internet and in Central New Jersey and Manhattan 

Back to top 


C6L1V@shaw.ca science forum Guru
Joined: 23 May 2005
Posts: 628

Posted: Wed Jul 19, 2006 9:47 pm Post subject:
Re: Expectation value in terms of cumulative distribution



Randy Poe wrote:
Quote:  I have to estimate some means and variances from empirical
data, and for various reasons what I have to work with are
estimates of cumulative distributions rather than densities
Actually, what I have are quantile levels, a set of values
q(p) = x such that P(X<=x) = F(x) = p at evenly spaced values of p
from 0 to 1, which gives me in effect estimates of F(x) at unevenly
spaced intervals of x.
So I started wondering if I could estimate E[x] and E[x^2] directly
from F(x) and came up with an elementary theorem in about
30 seconds which I don't remember seeing.

For any nonnegative r.v., EX = int_0^infinity G(x) dx, where G(x) =
P{X > x}. This is true, even if X does not have a density function f;
it is an old result that appears in just about every probability
textbook; see, eg., Ross, Introduction to Probability Models. (The
result holds in the sense that EX = infinity iff the integral diverges;
and if one side is finite, so is the other, and then they are equal.) I
guess you can reduce the [a,b] result to the >= 0 result by considering
X = a + Y, with 0 <= Y <= ba, and I think your E(X) expression follows
from that, whenever X has a density. However, I guess your actual
expression is new. If memory serves, there is a similar wellknown
expression for E(X^2), but my memory does not serve far enough to
retrieve the actual formula.
R.G. Vickson
Quote:  The proof is only
a couple of lines. Even though the proof seems pretty straightforward,
I'd like to know if this is in fact a theorem anyone has seen before.
Theorem: Let X be a random variable with finite support [a,b],
density function f(x) and cumulative probability F(x).
Then E[x] = b  integral(a,b)F(x) dx and
E[x^2] = b^2  2*integral(a,b) x F(x) dx
Proof: Integrate by parts.
E[x] = integral(a,b) x f(x) dx
Choose u = x, dv = f(x) dx => du = dx, v = F(x)
E[x] = (b*F(b)  a*F(a))  integral(a,b) F(x) dx which gives the first
result since F(a) = 0, F(b) = 1.
E[x^2] = integral(a,b) x^2 f(x) dx
= (b^2*F(b)  a^2*F(a))  integral(a,b) 2x F(x) dx
which gives the second result.
 Randy 


Back to top 


Ray Koopman science forum Guru Wannabe
Joined: 25 Mar 2005
Posts: 216

Posted: Wed Jul 19, 2006 11:12 pm Post subject:
Re: Expectation value in terms of cumulative distribution



Randy Poe wrote:
Quote:  I have to estimate some means and variances from empirical
data, and for various reasons what I have to work with are
estimates of cumulative distributions rather than densities
Actually, what I have are quantile levels, a set of values
q(p) = x such that P(X<=x) = F(x) = p at evenly spaced values of p
from 0 to 1, which gives me in effect estimates of F(x) at unevenly
spaced intervals of x.

If the pinterval is small enough, couldn't you estimate the first
and second raw moments by simple quadrature (e.g., trapezoids) on
q(p) and q(p)^2 ? 

Back to top 


ArtflDodgr science forum beginner
Joined: 09 May 2005
Posts: 45

Posted: Thu Jul 20, 2006 5:01 pm Post subject:
Re: Expectation value in terms of cumulative distribution



In article <1153344847.277486.315790@h48g2000cwc.googlegroups.com>,
"Randy Poe" <poespamtrap@yahoo.com> wrote:
Quote:  I have to estimate some means and variances from empirical
data, and for various reasons what I have to work with are
estimates of cumulative distributions rather than densities
Actually, what I have are quantile levels, a set of values
q(p) = x such that P(X<=x) = F(x) = p at evenly spaced values of p
from 0 to 1, which gives me in effect estimates of F(x) at unevenly
spaced intervals of x.
So I started wondering if I could estimate E[x] and E[x^2] directly
from F(x) and came up with an elementary theorem in about
30 seconds which I don't remember seeing. The proof is only
a couple of lines. Even though the proof seems pretty straightforward,
I'd like to know if this is in fact a theorem anyone has seen before.

It's well known. You can find the formula (valid for a nonnegative
random variable X)
E[X^b] = int_0^infty P[X>t] b*t^{b1} dt,
for b>0, on page 150 of the second edition of vol. II of Feller's
"An Introduction to Probability Theory and its Applications".
Without doubt, the formula predates the 1971 publication of that book by
many years. (For example, it appears as an exercise in Chung's
"Course", which first appeared in 1968; the inequality
E[X^b] <= int_0^infty P[X>t] b*t^{b1} dt
plays a role in Doob's 1953 proof of his L^p maximal inequality for
submartingales.)
Your formula results by applying the above formula to the r.v. X = xa.
Feller's proof is simply integration by parts, and yields the more
general expression (still for nonnegative X)
E[G(X)] = int_0^infty P[X>t] dG(t)
provided G: [0,infty) > [0,\inty) is nondecreasing.
Quote:  Theorem: Let X be a random variable with finite support [a,b],
density function f(x) and cumulative probability F(x).
Then E[x] = b  integral(a,b)F(x) dx and
E[x^2] = b^2  2*integral(a,b) x F(x) dx
Proof: Integrate by parts.
E[x] = integral(a,b) x f(x) dx
Choose u = x, dv = f(x) dx => du = dx, v = F(x)
E[x] = (b*F(b)  a*F(a))  integral(a,b) F(x) dx which gives the first
result since F(a) = 0, F(b) = 1.
E[x^2] = integral(a,b) x^2 f(x) dx
= (b^2*F(b)  a^2*F(a))  integral(a,b) 2x F(x) dx
which gives the second result.


A. 

Back to top 


Herman Rubin science forum Guru
Joined: 25 Mar 2005
Posts: 730

Posted: Thu Jul 20, 2006 5:03 pm Post subject:
Re: Expectation value in terms of cumulative distribution



In article <1153344847.277486.315790@h48g2000cwc.googlegroups.com>,
Randy Poe <poespamtrap@yahoo.com> wrote:
Quote:  I have to estimate some means and variances from empirical
data, and for various reasons what I have to work with are
estimates of cumulative distributions rather than densities
Actually, what I have are quantile levels, a set of values
q(p) = x such that P(X<=x) = F(x) = p at evenly spaced values of p
from 0 to 1, which gives me in effect estimates of F(x) at unevenly
spaced intervals of x.
So I started wondering if I could estimate E[x] and E[x^2] directly
from F(x) and came up with an elementary theorem in about
30 seconds which I don't remember seeing. The proof is only
a couple of lines. Even though the proof seems pretty straightforward,
I'd like to know if this is in fact a theorem anyone has seen before.

It is well known that if X is nonnegative with cdf F,
then E(X) = \int (1  F(t)) dt. It is less well known
that E(X) = \int^0 F(t) dt + \int_0 (1  F(t)) dt,
the first integral on the negative side, and the second
on the positive side. This can be extended to integrals
with arbitrary measures, and can even be used as a
definition of the LebesgueStieltjes integral.
Suppose X >= 0. This can easily be extended to the
expectation of X^2 which is \int P(X^2 > t) dt =
\int P(X > sqrt(t)) dt = \int (1  F(u))*2u du.
As you see, lots more can be done with this.

This address is for information only. I do not claim that these views
are those of the Statistics Department or of Purdue University.
Herman Rubin, Department of Statistics, Purdue University
hrubin@stat.purdue.edu Phone: (765)4946054 FAX: (765)4940558 

Back to top 


Randy Poe science forum Guru
Joined: 24 Mar 2005
Posts: 2485

Posted: Thu Jul 20, 2006 5:46 pm Post subject:
Re: Expectation value in terms of cumulative distribution



C6L1V@shaw.ca wrote:
Quote:  Randy Poe wrote:
I have to estimate some means and variances from empirical
data, and for various reasons what I have to work with are
estimates of cumulative distributions rather than densities
Actually, what I have are quantile levels, a set of values
q(p) = x such that P(X<=x) = F(x) = p at evenly spaced values of p
from 0 to 1, which gives me in effect estimates of F(x) at unevenly
spaced intervals of x.
So I started wondering if I could estimate E[x] and E[x^2] directly
from F(x) and came up with an elementary theorem in about
30 seconds which I don't remember seeing.
For any nonnegative r.v., EX = int_0^infinity G(x) dx, where G(x) =
P{X > x}.

OK. I did remember something like that for nonnegative rvs, but
couldn't remember the exact theorem and didn't have a
reference at hand. As I was scribbling this down and remembering
there was some sort of result with nonnegative rvs, I was
thinking  why do I need the nonnegative requirement?
Quote:  This is true, even if X does not have a density function f;
it is an old result that appears in just about every probability
textbook; see, eg., Ross, Introduction to Probability Models. (The
result holds in the sense that EX = infinity iff the integral diverges;
and if one side is finite, so is the other, and then they are equal.) I
guess you can reduce the [a,b] result to the >= 0 result by considering
X = a + Y,

Yes, I missed that trivial point. Others pointed out this connection
as well. I was thinking that this was, if not especially interesting,
at least different because it drops the nonnegativity requirement.
Oops.
So if I'm going to rival JSH for earthshaking new mathematics, I guess
it won't be on the basis of this little theorem.
Quote:  with 0 <= Y <= ba, and I think your E(X) expression follows
from that, whenever X has a density. However, I guess your actual
expression is new. If memory serves, there is a similar wellknown
expression for E(X^2), but my memory does not serve far enough to
retrieve the actual formula.

Since X^2 is nonnegative, it's probably a straightforward corollary.
Thanks all for the responses.
 Randy 

Back to top 


Google


Back to top 



The time now is Tue Nov 20, 2018 11:38 am  All times are GMT

