Return to Answer

As a segue to my prior post on this topic I want to share some tentative (albeit incomplete) exploration of the functions behind the linear algebra and related R functions. This is supposed to be a work in progress.

As a segue to my prior post on this topic I want to share some tentative (albeit incomplete) exploration of the functions behind the linear algebra and related R functions. This is supposed to be a work in progress.

House = function(A){
    Q = diag(nrow(A))
    reflectors = matrix(0,nrow=nrow(A),ncol=ncol(A))
    for(r in 1:(nrow(A) - 1)){ 
        # We will apply Householder to progressively the columns in A, decreasing 1 element at a time.
        x = A[r:nrow(A), r] 
        # We now get the vector v, starting with first entry = norm-2 of x[i] times 1
        # The sign is to avoid computational issues
        first = (sign(x[1]) * sqrt(sum(x^2))) +  x[1]
        # We get the rest of v, which is x unchanged, since e1 = [1, 0, 0, ..., 0]
        # We go the the last column / row, hence the if statement:
        v = if(length(x) > 1){c(first, x[2:length(x)])}else{v = c(first)}
        # Now we make the first entry unitary:
        w = v/first
        # Tau will be used in the Householder transform, so here it goes:
        t = as.numeric(t(w)%*%w) / 2
        # And the "reflectors" are stored as in the R qr()$qr function:
        reflectors[r: nrow(A), r] = w/t
        # The Householder tranformation is:
        I = diag(length(r:nrow(A)))
        H.transf = I - 1/t * (w %*% t(w))
        H_i  = diag(nrow(A))
        H_i[r:nrow(A),r:ncol(A)] = H.transf
        # And we apply the Householder reflection - we left multiply the entire A or Q
        A = H_i %*% A
        Q = H_i %*% Q
    }
    DECOMPOSITION = list("Q"= t(Q), "R"= round(A,7), 
            "contracted Q as in qr()$qr lower triang"=  
            ((A*upper.tri(A,diag=T))+(reflectors*lower.tri(reflectors,diag=F))), 
            "reflectors" = reflectors,
            "rho"=c(apply(reflectors[,1:(ncol(reflectors)- 1)], 2, 
                function(x) sum(x^2) / 2), ((-1)^nrow(A))*A[nrow(A),ncol(A)]))
    return(DECOMPOSITION)
}

   House = function(A){
    Q = diag(nrow(A))
    reflectors = matrix(0,nrow=nrow(A),ncol=ncol(A))
    for(r in 1:(nrow(A) - 1)){ 
        # We will apply Householder to progressively the columns in A, decreasing 1 element at a time.
        x = A[r:nrow(A), r] 
        # We now get the vector v, starting with first entry = norm-2 of x[i] times 1
        # The sign is to avoid computational issues
        first = (sign(x[1]) * sqrt(sum(x^2))) +  x[1]
        # We get the rest of v, which is x unchanged, since e1 = [1, 0, 0, ..., 0]
        # We go the the last column / row, hence the if statement:
        v = if(length(x) > 1){c(first, x[2:length(x)])}else{v = c(first)}
        # Now we make the first entry unitary:
        w = v/first
        # Tau will be used in the Householder transform, so here it goes:
        t = as.numeric(t(w)%*%w) / 2
        # And the "reflectors" are stored as in the R qr()$qr function:
        reflectors[r: nrow(A), r] = w/t
        # The Householder tranformation is:
        I = diag(length(r:nrow(A)))
        H.transf = I - 1/t * (w %*% t(w))
        H_i  = diag(nrow(A))
        H_i[r:nrow(A),r:ncol(A)] = H.transf
        # And we apply the Householder reflection - we left multiply the entire A or Q
        A = H_i %*% A
        Q = H_i %*% Q
    }
    DECOMPOSITION = list("Q"= t(Q), "R"= round(A,7), 
            "compact Q as in qr()$qr"=  
            ((A*upper.tri(A,diag=T))+(reflectors*lower.tri(reflectors,diag=F))), 
            "reflectors" = reflectors,
            "rho"=c(apply(reflectors[,1:(ncol(reflectors)- 1)], 2, 
                function(x) sum(x^2) / 2), A[nrow(A),ncol(A)]))
    return(DECOMPOSITION)
}

Now we are one degree away from the $\bf w$ vectors, and the first entry is no longer $1$, Hence the output of qr() will need to include the key to restore them since we insist on excluding the first entry of the "reflector" vectors to fit everything in qr()$qr. So are we seeing the $\tau$ values in the output? Well, no that would be predictable. Instead in the output of qr()$qraux(where this key is stored) we find $\rho=\sum \text{reflectors}^2/2=\bf \frac{w^Tw}{\tau^2} / 2$

Now we are one degree away from the $\bf w$ vectors, and the first entry is no longer $1$, Hence the output of qr() will need to include the key to restore them since we insist on excluding the first entry of the "reflector" vectors to fit everything in qr()$qr. So are we seeing the $\tau$ values in the output? Well, no that would be predictable. Instead in the output of qr()$qraux(where this key is stored) we find $\rho=\frac{\sum \text{reflectors}^2}{2}= \frac{\bf w^Tw}{\tau^2} / 2$.