I'm trying to find moments of of transformations of the multivariate normal distribution.
Let $x,y,z\sim N(\mu, \Sigma)$, I want to find out (for example) $E[e^x], E[e^{2x}], E[e^xy]$.
I have defined the following:
x, y, z = var('x y z')
X = vector([x,y,z])
mx, my, mz = var('mx my mz')
mu = vector([mx, my, mz])
sx, sy, sz = var('sx sy sz')
sxy, sxz, syz = var('sxy sxz syz')
assume(sx > 0)
assume(sy > 0)
assume(sz > 0)
S = matrix(SR, [[sx^2, sxy, sxz], [sxy, sy^2, syz], [sxz, syz, sz^2]])
f = (1 / sqrt(S.det() * (2*pi)^3))*exp(-0.5*((X - mu) * S.inverse() * (X - mu)))
And now I'm trying some integration, for example:
(x*f).integrate(z, -Infinity, Infinity).integrate(y, -Infinity, Infinity).integrate(x, -Infinity, Infinity)
I'd expect this to yield mx but of course life ain't easy so I get errors related to the covariance matrix:
ValueError: Computation failed since Maxima requested additional constraints; using the 'assume' command before evaluation *may* help (example of legal syntax is 'assume(sx^2*(sy^2*sz^2-syz^2)
>0)', see `assume?` for more details)
Is sx^2*(sy^2*sz^2-syz^2)
-sxy^2*sz^2
+2*sxy*sxz*syz
-sxz^2*sy^2 positive or negative?
Even for 1D integration (i.e only over $z$, (x*f).integrate(z, -Infinity, Infinity)) I get errors:
ValueError: Computation failed since Maxima requested additional constraints; using the 'assume' command before evaluation *may* help (example of legal syntax is 'assume(sx^2*(sy^2*sz^2-syz^2)
>0)', see `assume?` for more details)
Is sx^2*(sy^2*sz^2-syz^2)
-sxy^2*sz^2
+2*sxy*sxz*syz
-sxz^2*sy^2 positive or negative?
I have tried assuming everything he asks but I still get the same errors. Is there any nice way of telling SAGE to assume the matrix S is positive-semidefinite?
PS I have managed to work with two 2D case, by using correlation and the "full" version of the formula (rather than the vectorized):
x = var('x')
y = var('y')
mux = var('mux')
muy = var('muy')
sx = var('sx')
sy = var('sy')
r = var('r')
assume(r<=1)
assume(r>=-1)
k = 1/(2*pi*sx*sy*sqrt(1-r^2))
fxy = k*exp(-(1/(2*(1-r^2)))*(((x-mux)/sx)^2 - 2*r*((x-mux)/sx)*((y-muy)/sy) + ((y-muy)/sy)^2))
If I run integrate(integrate(x*fxy, y, -Infinity, Infinity), x, -Infinity, Infinity).simplify(), the result is eventually the correct one (mux).
