[R] regression constraints (again)

[Prof Brian Ripley]

It's easier to use crossprod, and you appear to have transposed A.

Here's a small example

library(quadprog)  # which you never mentioned!
X <- matrix(rnorm(100),, 4)
Y <- 0.25* X %*% rep(1, 4) + 0.1*rnorm(25)
D <- crossprod(X)
d <- crossprod(X, Y)
A <- cbind(1, diag(ncol(X)))
b <- c(1, rep(0,ncol(X)))
solve.QP(D, d, A, b, meq=1)

which works for me.

On Sun, 24 Aug 2003, Brett Robinson wrote:

> Im trying to do regressions with constraints that the weights 
> are all >=0 and sum(weights) = 1. I've read the archive and have 

weights?  I assume you mean coefficients?

> set the problem up with solve.QP and just the non-negativity constraints
> 
> along the lines of:
> 
> y as the data vector 
> X as the design matrix 
> 
> D <- t(X) %*% X 
> d <- t(t(y) %*% X) 
> A <- diag(ncol(X)) 
> b <- rep(0,ncol(X)) 
> fit <- solve.QP(D=D,d=d,A=t(A),b=b,meq=0) 
> 
> (as per Gardar Johannesson '01) 
> 
> When I try to add the extra constraint that sum(weights)=1 I get errors 
> owing to incompatibility of matrices. I add the constraint by putting an
> 
> extra column of all ones to A and setting meq=1.
> 
> I can work round it I think, by using an intercept and using the extra 
> column on the matrix for the sum(weights) constraint but I think that it
> should be possible without doing this.

> Grateful for any pointers as to where I am going wrong.

Please only send a message once, and do give the full details of what you
did so the helpers can spot your errors.  Also, don't appreviate argument
names: at best you confuse your readers, and you are indanger of confusing
yourself.

-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595