differential entropy of a gaussian

An important result is the differential entropy of a gaussian variable.

The pdf of a multivariate gaussian is

f(x)=1(2π)ndet(Σ)exp((xμ)TΣ1(xμ))

Now, we know that entropy is expectation of "surprise," or logp(x). "LOE" is "linearity of expectation."

H[X]=E[log(p(X))]Moved negative sign out, LOE=E[log(1(2π)n2det(Σ)12)12(xμ)TΣ1(xμ)]Substitution, square root to power=E[n2log(2π)+12log(det(Σ))12(xμ)TΣ1(xμ)]Log rule of multiplication=n2log(2π)+12log(det(Σ))+12E[(xμ)TΣ1(xμ)]LOE=n2log(2π)+12log(det(Σ))+12E[tr((xμ)TΣ1(xμ))]Term is a quadratic form resulting in a scalar, trace of scalar is itself=n2log(2π)+12log(det(Σ))+12E[tr(Σ1(xμ)(xμ)T)]commutation of trace=n2log(2π)+12log(det(Σ))+12tr(E[Σ1(xμ)(xμ)T])LOE=n2log(2π)+12log(det(Σ))+12tr(Σ1E[(xμ)(xμ)T])Covariance matrix is constant=n2log(2π)+12log(det(Σ))+12tr(Σ1Σ)Definition of covariance matrix=n2log(2π)+12log(det(Σ))+12tr(In)Matrix times its inverse is identity=n2log(2π)+12log(det(Σ))+12nTrace of identity is the length of diagonal=n2log(2πedet(Σ)1n)

Important result is that the differential entropy only depends on the variance of the gaussian, and not its mean.

Back to index