Ex. 3.11

Ex. 3.11

Show that the solution to the multivariate linear regression problem (3.40) is given by (3.39). What happens if the covariance matrices Σi are different for each observation?

Soln. 3.11

Like (3.38), we write (3.40) in matrix form

RSS(B;Σ)=tr[(YXB)TΣ1(YXB)].

By properties of trace operator, we have

RSS(B;Σ)=tr[(YTΣ1BTXTΣ1)(YXB)]=tr(YTΣ1YYTΣ1XBBTXTΣ1Y+BTXTΣ1XB).

Taking derivative and setting it to be zero, we get

RSS(B;Σ)B=XT(Σ1+(Σ1)T)XBXT(Σ1+(Σ1)T)Y(1)=0.

Note that Σ is a positive definite symmetric matrix, there exists S such that Σ1=SST. Therefore we obtain

B^=(XTSSTX)1XTSSTY=(XTSSTX)1XTSSTXXT(XXT)1Y=XT(XXT)1Y=(XTX)1XTXXT(XXT)1Y=(XTX)1XTY,

which is (3.39) in the text.

When Σi are different, the simple solution for B above does not hold. Instead, we have to deal with equations like (1) with different Σi. Numerical solutions are available though, as the problem is essentially in quadratic form of B.