Removing for loop in computing the finite difference Jacobian
11 vues (au cours des 30 derniers jours)
Afficher commentaires plus anciens
Hello friends,
I am trying to remove the for loop i.e. to vectorize the following code that computes the numerical Jacobian using finite difference. I think the vectorized version of the code will be faster than the current version especially for computing the Jacobian of a large-scale problem.
In addition, I will like to multiply J^T with v, where v is any vector with dimension equal to the rows of J, how can I include this in my code? Please, can anybody help me out?
if true
function [J]=numjac(f,x)
% computes the Jacobian of a function
tic
n=length(x);
fx=feval(f,x);
eps=1.e-8;
xperturb=x;
for i=1:n
xperturb(i)=xperturb(i)+eps;
J(:,i)=(feval(f,xperturb)-fx)/eps;
xperturb(i)=x(i);
end
toc
end
0 commentaires
Réponses (1)
Matt J
le 9 Jan 2018
Modifié(e) : Matt J
le 9 Jan 2018
I don't think you can vectorize further (for general functions f), but not pre-allocating J will definitely hurt you in large problems,
J=nan(length(fx),length(x)); %pre-allocate
for i=1:n
xperturb(i)=xperturb(i)+eps;
J(:,i)=(feval(f,xperturb)-fx)/eps;
xperturb(i)=x(i);
end
You could also use a PARFOR loop if you have the Parallel Computing Toolbox. Furthermore, the Optimization Toolbox least squares solvers have options to do all of the above for you internally.
4 commentaires
Matt J
le 9 Jan 2018
The efficient way to do that is always problem-specific. That's why Optimization Toolbox solvers like lsqnonlin give you the option of supplying a customized JacobianMultiplyFcn.
Voir également
Catégories
En savoir plus sur Performance and Memory dans Help Center et File Exchange
Produits
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!