mldivide versus least squares: X\(eye(m)) versus ( (X'X)\eye(m))*X'
5 vues (au cours des 30 derniers jours)
Afficher commentaires plus anciens
Dear all,
I am fitting a polynomial to data. I construct a polynomial basis X. I use some algorithm to update Y. I could use the mldivide to obtain coefficients theta or use
. But i don't know which one is more robust/accurate for my applicaiton. The system is normally overdetermined, but it might be exactly determined.

To obtain the coefficients of the polynomials I would normally do:
theta = X\Y;
However, since I have to do this repeatedly and X does not change I want to use:
%METHOD 1:
X_inv = X\eye(m);
%In each iteration:
theta = X_inv*Y;
where m is size(X,1).This should save computation time of the mldivide.
Now my questions is, for the Minimization of Squared Errors sometimes people also use
. In that case I should define:

%METHOD 2:
X_regr = ( (X'*X)\eye(m) )*X';
%In each iteration:
theta = X_regr*Y;
Should one method be preferred to the other (when overdetermined or exactly determined)? Or is that another method that is even better?
0 commentaires
Réponse acceptée
Matt J
le 6 Nov 2021
Modifié(e) : Matt J
le 6 Nov 2021
However, since I have to do this repeatedly and X does not change
If that's the case you should organize the different Y into the columns of a single matrix and do
X\[Y1,Y2,Y3,...,Yn]
Doing (X'*X)\ is not as numerically well-conditioned as X\, because the operation X'*X basically squares the condition number of X. Nevertheless, if X has many rows and few columns, X'*X\ will often run faster, and sometimes people will give priority to speed, especially if the cond(X) is known to be good.
3 commentaires
Plus de réponses (0)
Voir également
Catégories
En savoir plus sur Linear Algebra dans Help Center et File Exchange
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!