PROBABILITY DENSITY FUNCTION ESTIMATES 265

Equation (8.85),

V
"PW1«^ (· )
8 86

The constant c is dependent on the autocorrelation function of the data and the
sampling rate. For continuous bandwidth-limited white noise, experimental studies
indicate that c « 0.3. If the bandwidth-limited white noise is digitized to N=2BT
discrete values, the experimental studies indicate that c = 1.0, as would be expected
from the results of Equation (8.85).

8.3.3 Normalized rms Error

The total mean square error of the probability density estimate p(x) is the sum of
the variance defined in Equation (8.86) and the square of the bias defined in
Equation (8.82). That is,

c PJx)
2
W p"{x)
2

(p{x)-p{x)) 2
(8.87)
2BTW 24

Hence, the normalized mean square error is approximated by

W 4
J'{x)
ε \β{χ)
λ
(8.88)
2BTWp{x) 576 [p(x)

The square root gives the normalized rms error.

It is clear from Equation (8.88) that there are conflicting requirements on the
window width Win probability density measurements. On the one hand, a large value
of Wis desirable to reduce the random error. On the other hand, a small value of Wis
needed to suppress the bias error. However, the total error will approach zero as
Τ —> oo if W is restricted so that W —> oo and WT —> oo. In practice, values of
W< 0.2σ will usually limit the normalized bias error to less than 1%. This is true
χ

because of the p"(x) term in the bias portion of the error given by Equation (8.88),
Probability density functions of common (approximately Gaussian) random data do
not display abrupt or sharp peaks, which are indicative of a large second derivative.

8.3.4 Joint Probability Density Function Estimates

Joint probability density function estimates for a pair of sample time history records
x(t) and y(i) from two stationary (ergodic) random processes {x(t)} and (y(i)} may be
defined as follows. Analogous to Equation (8.66), let

P[ ,W ;y,W ]
X x y = ψ (8.89)

estimate the joint probability that x(t) is inside the interval W centered at x, while
x

simultaneously y(t) is inside the interval W centered at y. This is measured by the ratio
v

T y j T, where T represents the amount of time that these two events coincide in time
Xi x v
266 STATISTICAL ERRORS IN BASIC ESTIMATES

T. Clearly, T will usually be a function of both χ and y. This estimated joint

Xi>

probability will approach the true probability P[x, W ; y, W ] as Tapproaches infinity, x y

namely,

Ρ [χ, W ; y, W ] =
x y lim Ρ [χ, W ; y, W ] = x y lim ^ (8.90)

The joint probability density function p(x, y) is now defined by

p*,y)= hm . . . . . . — - = hm ———'-= hm p[x,y) (8.91)

_o
W i WxWy τ-^οο W Wy x

W ^Q y W,->0 W,-.0
Wy^O Wy->0

where

P[x,W ;y,Wy] J^y_

x =

^ W Wy X TW Wyx ^ '

Assume that W and W are sufficiently small that the bias errors are negligible.
x y

Then the mean square error associated with the estimate p(x, y) will be given by the
variance of the estimate. As for first-order probability density estimates, this quantity
is difficult to determine precisely by theoretical arguments alone. However, by using
the same heuristic arguments that produced Equation (8.86), a general form for the
variance can be approximated. Specifically, for the special case where x(t) and y (i) are
both bandwidth-limited white noise with identical bandwidfhs B,

Vai\p(x,y)] « ° ^ 2 p
(8.93)
\yy ,y)\ BTW W 2 x y
v ;

where c is an unknown constant.

8.4 CORRELATION FUNCTION ESTIMATES

Consider now two sample time history records x(t) and y(r) from two stationary
(ergodic) random processes {x{t)} and {y(r)}. The next statistical quantities of interest
are the stationary autocorrelation functions /?»(τ) and R (x) and the cross-correlation yy

function R^x). To simplify the following derivation, the mean values μ and μ will be χ ν

assumed to be zero. For continuous data, x(t) and y(r), which exist only over a time
interval Γ, the sample cross-correlation estimate Λ^,(τ) can be defined by

- * - f 1
x(t)y(t + x)dt 0<x<T
Τ-τ) 0

(8.94)
ι C t + x)dt -T<x<0
—γτ *(*Μ-