Random Signals for Engineers using MATLAB and Mathcad Copyright 1999 Springer Verlag NY Example 6.1 Correlation Functions for a Gaussian Process In this example we will compute the expected values of products of two random variables or the correlation coefficients for processes that have two-dimensional gaussian distributions. Specifically, we will evaluate E[X2Y2] and E[|X| |Y|] .The computations will make use of the results of Example 4.7. The equation for a two dimensional gaussian is syms x y rho sx sy mpi f=-1/2/(1-rho^2)*(x^2/sx^2-2*rho*x*y/sx/sy+y^2/sy^2); gauxy=1/2/mpi/sx/sy/(1-rho^2)^(1/2)*exp(f); pretty(gauxy) 2 2 x rho x y y --- - 2 ------- + --2 sx sy 2 sx sy exp(- 1/2 ---------------------) 2 1 - rho 1/2 -------------------------------2 1/2 mpi sx sy (1 - rho ) The conditioned gaussian density function will be needed or f ( y| X x ) f ( x, y) f ( x) In example 4.7 we found that fygx=-1/2/(1-rho^2)/sy^2*(y-rho*sy/sx)^2; gauygx=1/(2*mpi)^(1/2)/sy/(1-rho^2)^(1/2)*exp(fygx); pretty(gauygx) / rho sy\2 |y - ------| 1/2 \ sx / 2 exp(- 1/2 --------------) 2 2 (1 - rho ) sy 1/2 -----------------------------1/2 2 1/2 mpi sy (1 - rho ) We recognize that the conditional expression is just a gaussian density function in y with the mean E[Y | X x] and a variance of y x x 1 2 2 y The first correlation function that is needed E[X2Y2] and is computed using Equation 5.5-1 and is expressed as E[X2Y2] = E[X2E[Y2| X=x]] From the equation we recognize that we need to compute E[Y2 | X=x]. This can be computed directly from f(y | X= x) with a and b constants f ( y | x) 1 y b 2 exp 2 2 a 2a 1 Where a and b are b y x x and a y 1 2 The second moment is, using Matlab syms a b u agt=sym('a>0'); maple('assume',agt); EYgX=int(u^2*1/a*exp(-1/2/a^2*(u-b)^2),u,-inf,inf); pretty(EYgX) 2 b 1/2 2 after dividing by the normalization factor 1/2 pi 2 + a~ 1/2 2 1/2 pi 2 we obtain E[Y2|X=x}= a2+b2 It is instructive to trace the substitution used by Matlab to obtain this result. We must first make some substitutions to begin the process. Let us first make the u v substitution v=u-b , du = dv, and at the limits u v and u v syms v EYint=u^2*1/a*exp(-1/2/a^2*(u-b)^2); EYintv=subs(EYint,u,v+b); pretty(EYintv) 2 v (v + b) exp(- 1/2 ---) 2 a~ ----------------------a~ 2 Expanding the quadratic and expanding result in three integrals, the first and third have symmetrical kernels and can be integrated on the half interval and multiplying by 2 to obtain the result, the second integral is odd and will integrate to zero over the full interval. We perform the integration of the first integral using Matlab. EY1=1/a*exp(-1/2/a^2*v^2)*v^2; EYint1=2*int(EY1,v,0,inf); EYint1=simplify(EYint1); pretty(EYint1) 2 1/2 a~ 1/2 2 pi The integration of the third integral can be performed by observation . It is just the integral of the probability density function of a gaussian multiplied by b 2. Combining the result of the tree integral and normalizing by 2 we have E[Y | X=x] = a2 + b2 as we have shown above. Substitution for the constants, a and b, we obtain 2 E[Y | X x] 1 2 2 y 2 2 y2 x2 x2 Multiplying by x2 and taking the expected value we have y2 E X E [Y | X x ] 1 E X 4 x2 2 2 2 y 2 2 x 2 Where EX4 is E[X4]. We must now find this moment and we make use of the characteristic function for a gaussian in example 3.5. From Equation 3.3-7 we have after differentiating using Matlab and after substitution = 0 in the expression syms om sig EX4=real(i^4)*diff(exp(-om^2*sig^2/2),om,4) subs(EX4,om,0) EX4 = 3*sig^4*exp(-1/2*om^2*sig^2)-6*sig^6*om^2*exp(1/2*om^2*sig^2)+om^4*sig^8*exp(-1/2*om^2*sig^2) ans = 3*sig^4 Substitution in the expression we obtain the desired result. The result can be rewritten in more general terms as y2 EX Y EX E[Y | X x] 1 3 x4 2 x 2 2 2 2 2 y 2 2 x 2 syms sx sy EX2Y2=sy^2*(1-rho^2)*sx^2+rho^2*sy^2/sx^2*3*sx^4; pretty(simplify(EX2Y2)) 2 sy 2 sx 2 + 2 rho 2 sy 2 sx Interpretation in terms of the Expectation operation E[X 2Y2]= E[X2] E[Y2] + 2 E[XY]. The second correlation function, E[|X| |Y|], is computed using the integral definitions directly. Let us begin by evaluating and integral with an unnormalized the gaussian kernel in the first octant of the xy plane and we call this integral J. This integral cannot be evaluated directly by Matlab directly but after substitutions it can be evaluated. iarg=exp(-u^2-v^2-2*rho*u*v); pretty(iarg) int(int(iarg,u,0,inf),v,0,inf) 2 2 exp(-u - v - 2 rho u v) ??? Error using ==> sym/sym (char2sym) Definite integration: Can't determine if the integral is convergent.Need to know the sign of --> 2*rho*vWill now try indefinite integration and then take limits.1/2*pi^(1/2)*exp(v^2*(rho-1)*(rho+1))1/2*erf(rho*v)*pi^(1/2)*exp(v^2*(rho-1)*(rho+1)) is not a valid symbolic expression. Using a linear transformation from u, v to p, q we transform the kernel syms p q u=p-rho/(1-rho^2)^(1/2)*q; v=q/(1-rho^2)^(1/2); iarg=(-u^2-v^2-2*rho*u*v); pretty(simplify(iarg)) 2 -p 2 - q The Jacobian is jac=jacobian([u;v], [p; q]) jac = [ [ 1, -rho/(1-rho^2)^(1/2)] 0, 1/(1-rho^2)^(1/2)] det(jac) ans = 1/(1-rho^2)^(1/2) Substitution for p and q into the kernel, changing the derivative terms and at the limits we have jint=1/(1-rho^2)^(1/2)*int(int(exp(-p^2-q^2),p,q*rho/(1rho^2)^(1/2) \ ,inf),q,0,inf); pretty(jint) rho 1/4 pi - 1/2 atan(-------------) 2 1/2 (1 - rho ) -------------------------------2 1/2 (1 - rho ) The term can be replaced in the above equation a tan 1 2 2 a cot 1 2 and the integral becomes a cot 1 2 2 1 2 1 J 1 2 1 2 We have expressed the acot term in terms of and angle . Using trigonometry we can form a right triangle with the side opposite the angle = (1- 2)1/2 and the side adjacent equal to and the hypotenuse equal to 1. The above expression can be reduced to J where cos 1 2 1 2 a cos 1 csc( ) 2 The above expression can be differentiated to obtain the left-hand side syms u v syms phi iarg= -u^2-v^2-2*cos(phi)*u*v; iargd=diff(iarg,phi) iargd = 2*sin(phi)*u*v and the right hand side jrs=1/2*phi*csc(phi); jrsd=diff(jrs,phi) jrsd = 1/2*csc(phi)-1/2*phi*csc(phi)*cot(phi) Combining the terms in from the left half of the equation above we obtain the normalized integral equation evaluation u v e 0 u 2 v 2 2uvcos du dv 0 1 2 csc 1 cot 4 The E[|X| |Y|] can be expressed in terms of the above equation . First the expression is written. This result must be normalized because this expected value is 4 times the expected value of the gaussian in the first octant. This is due to the absolute value operators in E[|X| |Y|] . Using the equation for a gaussian density function with x2 y2 2 and, for simplicity, we have taken to be negative and used the symbol EXY is used for E[|X| |Y|] in the expressions we have. EXY 4 2 2 1 2 0 0 x ye x 2 y 2 2 x y cos dx dy This integral can be evaluated by rewriting in term of the integral developed above. This variable scaling of the argument changes the form of the integral to the one evaluated above. We have for the argument of the kernel syms x y x=u*sig*2^(1/2)*(1-rho^2)^(1/2); y=v*sig*2^(1/2)*(1-rho^2)^(1/2); arguv=-1/2/(1-rho^2)/sig^2*(x^2+y^2+2*rho*x*y); pretty(simplify(arguv)) 2 -u 2 - v - 2 rho u v and the differentials 4 2 1 2 2 dx dy 1 du dv 1 2 We now obtain the integral EXY 42 2 1 2 3 2 0 0 u ve u 2 v 2 2 x ycos du dv From above, this integral can be evaluated EXY 42 3 1 2 2 sin csc 1 cot 4 Expanding and simplifying we obtain EXY=4*2/mpi*sig^2*(sin(phi))^3*(1/4*(csc(phi))^2*(1-phi*cot(phi))); pretty(simplify(EXY)) 2 sig (-sin(phi) + phi cos(phi)) -2 ------------------------------mpi to the desired result The integral can also be expressed in term of the complement of , or + = /2 and we have EXY 2 2 cos sin This equation is used in Example 6.2 to evaluate the correlation function.